dissect.dev
Malicious sev:medium family:InstallUnion 2026-11-17

ByteBreaker

Lure claim: “"ByteBreaker V1" — a free Roblox script executor / exploit, advertised by its own publisher description as having "no key system" ("the ads shown in the installer help this exploit"). The file masquerades as the product download while declaring itself FileDescription/ProductName = Downloader, Compan”

Specimen Hashes
SHA-256bd1f1619f850be8c84ae0acdd0bb7ff3d88b8cb5647877c078d7805d7a55582f
SHA-1932e7cce8f6cde321f2288ccaaeb8ae5d84ab26d
MD5280603b04aeadbb0d42f6043354923cd
Imphashafb4e05a0cdf45e3ff8be8f3da3dd155
FilenameByteBreaker V1_85892839.exe
Executable Profile
TargetAMD64 (x86-64), PE32+, Windows GUI subsystem (IMAGE_SUBSYSTEM_WINDOWS_GUI), nati · Windows GUI
EP / BaseEP 0x00061400 · Base 0x140000000
Size / Entropy756.9 KB · 6.42
PackerNone
COFF Time2026-10-29T19:51:04Z
Originhttps://bytebreaker.cc/
Evasion Tactics & Structural Indicators
  • T1204.002 User Execution: Malicious File - user double-clicks <Product> V1_<8-digit>.exe
  • T1036.005 Masquerading: Match Legitimate Name/Location - %TEMP%\ADR<hex4>.exe and GLY<hex4>.exe via GetTempFileNameW + DeleteFileW + ".exe"
  • | 248 | | 16 AEAD envelope, zstd inside, runtime-derived key
  • T1518 Software Discovery - per-offer registryChecks (RegGetValueW) and fileChecks (GetFileAttributesW) gate presentation
  • T1071.001 Application Layer Protocol: Web Protocols - WinINet POST https://dlsft.com/service.php, UA exactly "Mozilla/5.0", INTERNET_OPEN_TYPE_DIRECT (ignores system proxy)
  • T1105 - URLDownloadToFileW of 16.7 MB archive over plain HTTP; URLOpenBlockingStreamA for in-memory creatives
  • T1548 Abuse Elevation Control Mechanism - manifest requireAdministrator (requireAdministrator uiAccess=false) taken before offer set is known; all children inherit token
  • T1496 - three of five offers are proxyware: idle uplink becomes residential proxy egress (Honeygain SDK 1.5.0 signed Honeygain UAB; "Your device may function as a network exit node")
  • T1562.001 Disable or Modify Tools - publisher FAQ instructs distributors: "recommend your audience to disable Windows Defender and any Anti-Virus during the download process" (advisory, not automated)
  • T1204.001 Malicious Link - affiliate redirect https://www.ovardu.com/6R9H6BD/KMZXBTT/?sub1=FileDM&sub=default for CPA browser install
  • PE Anomaly Appended PE overlay (8,072 bytes, entropy 7.48)
  • PE Anomaly Authenticode signature table present (verify certificate chain / stolen cert)
PE Section Table (7)
SectionVirt AddrVirt SizeRaw SizePermsEntropy
.text 0x00001000 586,604 B 586,752 B R-X 6.44
.rdata 0x00091000 138,062 B 138,240 B R-- 5.22
.data 0x000b3000 16,556 B 10,240 B RW- 3.64
.pdata 0x000b8000 23,772 B 24,064 B R-- 5.84
.fptable 0x000be000 256 B 512 B RW- 0.00
.rsrc 0x000bf000 1,568 B 2,048 B R-- 3.94
.reloc 0x000c0000 3,780 B 4,096 B R-- 5.31

ByteBreaker V1_85892839.exe is a 775,048-byte signed x64 GUI download wrapper, not the Roblox exploit it appears to be. It carries no product name, icon, dialog or string-table resources and under 5 KB of application literals; its entire visible interface (five pages, up to five offers, legal text, banners) is supplied at run time by https://dlsft.com/service.php. Two synchronous POSTs drive it — request=offers (bundle, keyed on a channel) and request=getDownloadInfo&name=<stem> (product descriptor, keyed on the trailing eight-digit identifier only, confirmed by 701 sequential probes and alias testing; WhateverName_85892839 resolves identically). The .exe suffix is therefore a server-side row key, and the "product" is one row in a mutable table.

Retrieval is obfuscated by a home-grown hotlink mask: s = %08x(crc32(id XOR "SCI_multipart")), without which directfiledl.com/getfile answers Security code is not matching. The 16.7 MB result is a third-party Luau executor, Xeno 1.4.816 (CodeView path C:\Users\rizve\OneDrive\Desktop\Projects\Xeno\build\Release\net9.0-windows8.0\Xeno.pdb), re-branded with nine strings by a repackager calling itself Evon; the engine still says "Xeno" 67 times and "ByteBreaker" zero. It is NativeAOT .NET 9 + statically linked Luau, exports Attach/Execute/GetClients/Initialize/SetSetting/Version, embeds cpp-httplib on loopback, Discord RPC with join secrets, and pulls its bytecode-patch configuration from an AEAD-sealed 276-byte blob mirrored on Pastebin and Cloudflare Pages (key derived at runtime; 9.37 M static candidates excluded).

Monetisation is bandwidth, not data: three of five offers are proxyware (two first-party, one shipping Authenticode-signed hgsdk.dll 1.5.0 from Honeygain UAB, Vilnius, connect.sdk.honeygain.com), one is a CPA browser redirect, one is the comScore/VoiceFive panel with special-category consent text. No component reads browser credential stores, cookies, the Roblox security cookie, wallets or hardware identifiers; telemetry is a server-issued 64-bit install id, an IP-derived country, a consent version and eight funnel events, one of which travels over cleartext HTTP. The real exposures are structural: Administrator acquired before the offer set is known (no offer requires machine scope), product integrity unverified over plain HTTP from a single Hetzner host with no TLS listener, and a stale packaged engine (1.2.0 vs manifest 1.4.816) that self-updates from an authenticated-by-hash-only manifest.


Static Reverse Engineering Assessment

ByteBreaker V1_85892839.exe and Its Delivery Chain


Report ID RE-2026-1117-BB
Version 1.0 (final)
Report date 17 November 2026, 02:10
Analysis period 16–17 November 2026
Classification TLP:CLEAR
Subject ByteBreaker V1_85892839.exe (775,048 B)
Associated artefacts 6 (see Table 2.1)
Analyst Elera
Method Static PE/C++ analysis, managed-metadata recovery, ASN.1 signature analysis, live endpoint interrogation, catalogue enumeration, cryptographic key-space sweep, OSINT corroboration
Companion file ByteBreaker_V1_85892839_ANALYSIS.md (working notes, 3 301 lines)

Revision history

Ver Date Change
0.1 2026‑11‑16 Initial triage; PE geometry, signature, literal pool
0.9 2026‑11‑17 Control-flow recovery; protocol reconstruction; payload acquisition and identification
1.0 2026‑11‑17 Offer-bundle analysis; third-party SDK identification; commercial attribution; risk model; indicators; corrections log

1. Executive summary

The Subject is a 775 KB, 64-bit Windows GUI executable that presents itself as a download manager for a Roblox script executor. Analysis establishes that it is a white-label pay-per-install (PPI) distribution wrapper, built by the Dutch software and file-hosting company AN Soft (trading as FileDM) and operated inside the InstallUnion advertising network as campaign 1538. The wrapper contains no product code. At execution it contacts two provider endpoints, retrieves a product descriptor keyed on its own filename, downloads a 16.7 MB archive over cleartext HTTP, and — subject to a locally evaluated targeting model — offers the user up to five additional software installations, four of which it executes silently with inherited Administrator rights.

The delivered archive is Xeno 1.4.816, a free Luau-based Roblox executor authored by an independent developer using the handle rizve, re-branded by a third party using the name "Evon". The re-brand required four HTML strings, one version resource and two file renames; every other product string in the archive still identifies Xeno.

The offer bundle, not the executor, is the commercially significant component. Three of the five offers are bandwidth-sharing clients: applications that, once accepted, sell the host's idle upload capacity into a residential proxy network. One of them ships the Honeygain SDK 1.5.0 (hgsdk.dll), an Authenticode-signed component from Honeygain, UAB of Vilnius, Lithuania. A second ships WinDivert, a kernel-mode packet interception driver, alongside a packet-level advertisement blocker. The remaining two offers are a mainstream browser distributed through a cost-per-action affiliate network and a comScore market-research panel whose acceptance text carries the most extensive data scope in the bundle.

No component in the chain was observed to read browser credential stores, browser cookies, the Roblox security cookie, cryptogram wallet files, hardware identifiers, or the operating-system username. The Subject's own telemetry is limited to a server-issued installation identifier, an IP-derived country code, a consent version, and nine funnel events. The privacy-relevant behaviour of the whole chain is therefore concentrated in the offers and in the third-party SDK they embed, not in the Subject itself.

Nine key judgements are recorded in §1.1 with associated confidence and severity ratings.

1.1 Key judgements

ID Judgement Confidence Severity
KJ‑01 The Subject is a generic, multi-purpose PPI wrapper — not bespoke software for the delivered product. It carries no product name, icon, dialog template, string table or payload. High Informational
KJ‑02 The Subject is delivered signed with an organisation-validated (OV) ECDSA code-signing certificate to a one-developer Dutch entity, AN Soft, and is therefore an authentic artefact of that vendor rather than a tampered copy. High Informational
KJ‑03 The delivered product is a third-party re-brand of the free executor Xeno by rizve. Requesting the descriptor under ByteBreaker V1_85892839 returns ByteBreaker V1 / bytebreakerV1.zip; the archive contains zero occurrences of the string "ByteBreaker" outside a 28 KB HTML file and one version resource. High Low
KJ‑04 Product identity is resolved server-side, by numeric identifier, not locally: any name with the suffix _85892839 resolves to the same product, and no other identifier in a 701-value sample resolves. The identifier is a provider database key. High Informational
KJ‑05 Monetisation is effected by bandwidth resale, not by advertising, subscription or credential theft: three of five offers are clients for residential-proxy networks, and two of those are first-party products of the same vendor that supplied the wrapper. High Moderate
KJ‑06 The Honeygain SDK 1.5.0 (hgsdk.dll, Authenticode: CN=Honeygain, UAB, SERIALNUMBER=306103177, Vilnius LT) is embedded in one offer. The SDK exposes an explicit consent API (hgsdk_request_consent, hgsdk_opt_in, hgsdk_opt_out, hgsdk_is_opted_in); the integration's consent path was not resolved statically and is rated an open question. High (presence) / Moderate (consent flow) Moderate
KJ‑07 The Subject acquires Administrator rights before it knows which offers it will present, and passes that token to five payloads, four of which run silently. None of the five offers requires machine-wide installation; one explicitly requests per-user installation (--allusers=0). Elevation is therefore speculative, not functional. High Moderate‑High
KJ‑08 Both the product descriptor and the product archive are retrieved at run time and are mutable server-side. The archive is also stale on delivery (bin/current_version.txt = 1.2.0 against version.json = 1.4.816), indicating a further update hop after first launch. Static assessment of the client therefore does not transitively assess the delivered code at a later date. High Moderate
KJ‑09 The product archive is served from directfiledl.com over plain HTTP, from a single Hetzner host with no TLS listener on 443, and is not signature-verified by the Subject. Integrity of the delivered product rests on a constant embedded in the Subject (§5.4). High Moderate

1.2 Bottom line

The Subject is a correctly built, professionally signed, competently written installer whose purpose is not the software it installs. It is a monetisation shell: a generic wrapper that obtains elevation speculatively, resolves what to offer from a mutable server-side model, and converts the user's installation event into revenue by reselling their upload capacity. Its engineering quality is higher than its transparency. The residual risk in the chain does not arise from concealment of what it reads — an analysis of every component found no credential-collection primitives — but from the aggregate of four silent, Administrator-level installers, two of which are proxyware and one of which loads a kernel driver, selected by a factor the recipient cannot see.


2. Scope, method and limitations

2.1 Objects of analysis

ID Object Size Status
S‑00 ByteBreaker V1_85892839.exe 775,048 B Primary subject; complete static analysis
S‑01 product descriptor (JSON, service.php) 232 B Retrieved from remote
S‑02 offer bundle (JSON, service.php) 8 887 B Retrieved from remote (5 offers)
S‑03 bytebreakerV1.zip (delivered product archive) 16,787,263 B Retrieved; fully extracted
S‑04 Glory-1.0.7-windows-x64.zip (offer 2) 12,579,642 B Retrieved; fully extracted
S‑05 AdReject-1.0.9-windows-x64.zip (offer 3) 574,773 B Retrieved; fully extracted
S‑06 aichat_2.0.3_x64-setup.exe (offer 4) 3,807,352 B Retrieved; unpacked (NSIS)
S‑07 ContentI3.exe reference (offer 5) – Referenced from feed; not retrieved

Artefacts were retrieved directly from the vendor's public endpoints during the analysis window. Retrieval was performed with ordinary GET requests except where the published protocol requires POST (§5.2); no authentication, no registration and no client-specific state was used.

2.2 Techniques applied

  1. Static PE analysis (headers, sections, directories, entropy, relocations, resources, overlay accounting).
  2. Signature analysis: PKCS#7 parsing, certificate path recovery, EKU, validity, RFC‑3161 counter-signature extraction.
  3. Build provenance: CodeView RSDS record recovery, Rich-header decoding, linker section-table extraction from the debug directory.
  4. Full-function decompilation (1 654 functions) and manual control-flow recovery of the 28 KB application layer.
  5. Literal-pool extraction across both encodings (ASCII/UTF‑16), with RVA attribution and cross-referencing into code.
  6. Managed-metadata recovery (BSJB, TypeDef/MethodDef tables, AssemblyRef) for the .NET components.
  7. Resource-tree enumeration and version-resource decoding.
  8. Protocol reconstruction by reading format-string call sites, then live validation of every endpoint retained by the analysis.
  9. Catalogue enumeration: 701 sequential identifier probes with type-safe response handling.
  10. Cryptographic envelope inference by length decomposition, followed by a 9 366 416-candidate key-space sweep over all byte offsets and both AES key sizes.
  11. OSINT corroboration of the upstream product identity and the vendor's claims.
  12. Cross-component corroboration of a single campaign identifier across three independent codebases (§7.4).

2.3 Confidence convention

Term Meaning
High Multiple independent artefacts agree; would require a deliberate, difficult forgery to contradict
Moderate One strong artefact plus consistent circumstance; a specific hypothesis remains unfalsified
Low Inference from pattern or convention only; alternative explanations exist
Open Experimentally reachable but not reached within this engagement; method is specified

2.4 Limitations

  • Static and remote-probe analysis only. The Subject and its payloads were not executed for behavioural capture. No sandbox/ETW trace, no Procmon, no packet capture, no dynamic debugging session was performed. Statements about runtime behaviour are derived from code, literals, imports and call structure, and are corroborated against the vendor's own served documents where possible.
  • No environment stratification. All remote interrogation (countries, offers, catalogue) was performed from a single egress location in the United Kingdom, between 16 and 17 November 2026, while the feed reports itself as varying "per country and day". Country-dependent behaviour other than that returned for GB is inferred, not measured.
  • Point-in-time feeds. The offer bundle and product descriptor are mutable. Only offer id values and the provider's numeric identifier are stable keys.
  • One configuration of the delivered product. The packaged engine was stale against its manifest (§6.6); the post-update engine was not obtained.
  • The encrypted remote configuration was not opened. The envelope and its role are established; the plaintext is not (§6.7).
  • No disablement test. Determinations of "not present" rest on absence of literals, imports and call sites. For the components that were fully decompilable this is strong; for the Themida-protected launcher it is weaker and is stated as such wherever it applies.

2.5 Terminology

PPI pay-per-install · PUP/PUA potentially unwanted program/application · OV/EV organisation- validated/extended-validation (code signing) · AoB artefact of build · NativeAOT ahead-of-time native compilation of a .NET assembly · proxyware software that supplies its host's bandwidth as a residential proxy egress · NativeAOT.


3. Technical identification

3.1 Primary sample

Property Value
File name ByteBreaker V1_85892839.exe
Size 775,048 bytes (0xBD388)
Format PE32+ executable, AMD64, GUI subsystem, 7 sections
SHA‑256 bd1f1619f850be8c84ae0acdd0bb7ff3d88b8cb5647877c078d7805d7a55582f
SHA‑1 932e7cce8f6cde321f2288ccaaeb8ae5d84ab26d
MD5 280603b04aeadbb0d42f6043354923cd
Authentihash (SHA‑256) 47cb2bce143f52236604570e76081fa7bbf49c27955cba8be9ea4780f1d1b1c1
imphash b3eb077da70cf3699e3dc8149d29c2be
Image base / entry point 0x140000000 / 0x140061400
SizeOfImage 0xC1000
Link timestamp (UTC) 2026‑10‑29 19:51:04
PE checksum 0xC6DED — equals computed value
Overlay 8,072 B = Authenticode certificate only
Signature Present, valid, countersigned (§3.4)

The PE checksum matching the calculated value, and the overlay consisting solely of the certificate table, are consistent with a stock linker output that has not passed through a post-link patcher, binder, or resource editor.

3.2 Section table

Section VA Virtual size Raw ptr Raw size Entropy Characteristics
.text 0x1000 586,604 0x400 586,752 6.442 CODE | EXEC | READ
.rdata 0x91000 138,062 0x8F800 138,240 5.222 IDATA | READ
.data 0xB3000 16,556 0xB1400 10,240 3.641 IDATA | READ | WRITE
.pdata 0xB8000 23,772 0xB3C00 24,064 5.840 IDATA | READ
.fptable 0xBE000 256 0xB9A00 512 0.000 IDATA | READ | WRITE
.rsrc 0xBF000 1,568 0xB9C00 2,048 3.937 IDATA | READ
.reloc 0xC0000 3,780 0xBA400 4,096 5.309 IDATA | READ, DISCARDABLE

Observations:

  • .text is not writable (R-X) and .text entropy is 6.44 — consistent with ordinary optimised code, not with a runtime unpacking stage, which requires either a writable code section or an explicit VirtualProtect transition. VirtualProtect is imported but is consumed by the UCRT.
  • .fptable is entirely zero, 160 bytes of contribution. It is a linker-generated function-pointer table (VS 2019 16.x / VS 2022 link.exe), not a protector section. Its presence is one of several independent indicators of the toolchain generation (§3.5).
  • Section raw offsets are strictly sequential with no gaps between them (0x400 + 586,752 = 0x8F800, and so on), and the final raw section ends at 0xBB400 (766,976), which is exactly the start of the certificate table. There is no unaccounted byte in the file.
  • No TLS directory, no delay-load directory, no bound-import directory, no CLR descriptor.

3.3 Resource tree

Only two resource types are present:

RT_VERSION   (0x10)  id 1  lang 0x0409/0x04B0   600 B
RT_MANIFEST  (0x18)  id 1  lang 0x0409          808 B

There is no RT_ICON, RT_GROUP_ICON, RT_DIALOG, RT_STRING, RT_BITMAP, RT_MENU, RT_ACCELERATOR, RT_HTML or RT_TOOLBAR. Three consequences follow, and they are load-bearing for later sections:

  1. No icon resource ⇒ the file displays with the generic default application icon in Explorer and in the task bar, and carries no visual identity.
  2. No dialog or string-table resources ⇒ every control, label and message is instantiated or supplied at run time. The user interface cannot be enumerated statically from resources; it must be recovered from code (§4.3), and the text it displays is not present in the binary at all.
  3. No product identity ⇒ there is nothing in the Subject, at any layer, that names the software it distributes. This is the design premise of the whole class (§3.6, §6.4).

3.4 Version resource and manifest

CompanyName      AN Soft
FileDescription  Downloader
FileVersion      1.0.0.0
InternalName     Downloader
LegalCopyright   © AN Soft
OriginalFilename (absent)
ProductName      Downloader
ProductVersion   1.0.0.0
Translation      0x0409 / 0x04B0   (US English, Unicode)
VS_FIXEDFILEINFO 0xFEEF04BD, FileOS=4 (NT), FileType=1 (application), FileSubtype=0

The StringFileInfo block contains no CompanyWebsite, Comments, PrivateBuild or SpecialBuild key — fields that are present in the overwhelming majority of commercially produced executables and are the normal place for a support URL or an e-mail address.

The embedded manifest declares, in full:

<assembly …>
  <dependency><dependentAssembly>
    <assemblyIdentity type="win32" name="Microsoft.Windows.Common-Controls"
      version="6.0.0.0" processorArchitecture="*"
      publicKeyToken="6595b64144ccf1df" language="*"/>
  </dependentAssembly></dependency>
  <trustInfo><security><requestedPrivileges>
    <requestedExecutionLevel level="requireAdministrator" uiAccess="false"/>
  </requestedPrivileges></security></trustInfo>
  <application><windowsSettings>
    <dpiAware xmlns="http://schemas.microsoft.com/SMI/2005/WindowsSettings">True/PM</dpiAware>
  </windowsSettings></application>
</assembly>

Two entries are operationally significant and are revisited in §4.5 and §10.3:

  • requestedExecutionLevel level="requireAdministrator" — unconditional elevation at every start, with a UAC prompt, before any user interface exists and before the software knows what it will install (§1.1 KJ‑07).
  • dpiAware = True/PM — declares per-monitor DPI awareness (the v1 True/PM form) but no dpiAwareness/gdiScaling fallback, and the code contains no AdjustWindowRectExForDpi or per-monitor scaling logic. On mixed-DPI systems the owner-drawn interface therefore renders at system scale and is not rescaled on monitor change.

<compatibility> (supportedOS GUIDs) is absent, so Windows applies no version lie and the process is not declared Windows 10/11 aware.

3.5 Build provenance

3.5.1 Debug directory

Four IMAGE_DEBUG_DIRECTORY entries (RVA 0xA35B0, 112 bytes total, all sharing the same TimeDateStamp):

# Type Size Content
0 2 CODEVIEW 83 RSDS — CodeView record (§3.5.2)
1 12 20 linker-specific
2 13 940 GCTL — COFF section table (§3.5.3)
3 14 0 empty

3.5.2 CodeView record

Signature : RSDS
GUID      : 861b2640-7ce5-cb4a-bc2f-99184c3d8202    (bytes 861b26407ce5cb4abc2f99184c3d8202)
Age       : 3
PDB       : C:\Users\Samim\Desktop\Installer\x64\Release\Installer.pdb

This single record supplies four pieces of analytic information, listed in §3.6. The build was made without path normalisation (/pathmap), without deterministic source-path stripping, and without a project-name rename — i.e. from a Visual Studio default configuration on a workstation desktop.

3.5.3 Linker section table (GCTL, debug type 13)

The table enumerates the COFF section contributions the executable was assembled from:

.text$di  .text$mn  .text$mn$00  .text$mn$21  .text$x  .text$yd
.idata$2  .idata$3  .idata$4  .idata$5  .idata$6
.00cfg  .CRT$XCA  .CRT$XCAA  .CRT$XCC  .CRT$XCL  .CRT$XCU  .CRT$XCZ
.CRT$XIA  .CRT$XIAA  .CRT$XIAC  .CRT$XIC  .CRT$XIZ
.CRT$XPA  .CRT$XPX  .CRT$XPXA  .CRT$XPZ  .CRT$XTA  .CRT$XTZ
.rdata  .rdata$00  .rdata$r  .rdata$voltmd  .rdata$zzzdbg
.rtc$IAA  .rtc$IZZ  .rtc$TAA  .rtc$TZZ
.xdata  .xdata$x  .bss  .fptable  .rsrc$01  .rsrc$02
Fragment Inference
.CRT$XCAA…$XTZ, .CRT$XPA…$XPZ complete UCRT initialiser chains ⇒ statically linked CRT (/MT)
.rtc$IAA/.IZZ/.TAA/.TZZ, .rdata$r RTTI runtime-check data and type descriptors ⇒ /GR enabled
.rdata$voltmd volatile-metadata table — emitted only by VS 2019 16.x and later linkers
.text$mn, .text$mn$00, .text$mn$21 COMDAT//Gy bucket placement under /GL
.text$di, .text$yd, .bss dynamic initialisers, atexit destructors, zero-init globals ⇒ C++ with non-trivial globals
.fptable x64 function-pointer table (see §3.2)
.00cfg __guard_check_icall_fptr boilerplate

3.5.4 Rich header

Present at file offset 0x3C (DanS), terminator Rich at 0xF8, XOR key 0x9176119f. The three decoded padding dwords (0x0EBA1F0E, 0x0CD09B400, 0x04C01B821) are the standard MSVC /GS and error-report helper markers; the associated __security_cookie (0x1400b3180) is referenced in the prologue of every non-trivial function, confirming /GS was enabled project-wide. The header is intact and internally consistent — no post-link editing.

3.5.5 Toolchain determination

Evidence Conclusion
No VCRUNTIME140/MSVCP140/api-ms-win-crt-* in the import table, full api-ms-win-crt-* string set present statically linked UCRT (/MT)
.CRT$X* chains, .bisect-free layout, MSVC mangling throughout Microsoft Visual C++
.rdata$voltmd, .fptable, blank-name sections tolerated linker VS 2019 16.x or VS 2022
nlohmann banner …3.11.2 (released 2022‑08) source tree not older than 2022 Q3
std::filesystem error strings ("state not recoverable", "owner dead", "cross device link") STL VS 2015 U3 or later, practically C++17
std::wstring_convert<codecvt_utf8_utf16<wchar_t,1114111,0>> RTTI (deprecated in C++17, removed in C++26) C++17 dialect, adherence to the pre-u8 era
DllCharacteristics 0x8160 — no GUARD_CF bit /guard:cf not used
Prologue cookie loads /GS on
GCTL present /DEBUG on (so a PDB exists upstream)

Determination: MSVC, x64, Release, /O2 /GL /LTCG, /MT, C++17, RTTI on, /GS on, Control Flow Guard off. Consistent with a default Release|x64 configuration of Visual Studio 2019 or 2022, in which the developer changed nothing except the source.

3.6 What the provenance record discloses

The CodeView record is the highest-value artefact in the Subject, because unlike code it cannot be written to lie without deliberate effort.

Disclosed Value Analytic use
Windows profile C:\Users\Samim developer account name; corroborates the signature subject
Project name Installer the product is called "Installer"; also the window class, window title and .text section name
Configuration x64\Release default VS wizard output; no custom output directory
Build locality …\Desktop\… built from the desktop, not from a repository working copy, CI agent, or mapped drive
Symbol name Installer.pdb project file Installer.vcxproj (or equivalent)

The absence of any repository-derived path (\src\, \git\, \build\, a solution folder) and the presence of OneDrive/Desktop in the sibling products (§8) together indicate a single-developer, non-CI, desktop-based build process across the whole product line.

3.7 Authenticode

3.7.1 Structure

WIN_CERTIFICATE at file offset 0xBB400, dwLength = 0x1F88 (8 072), wRevision = 0x0200, wCertificateType = 0x0002. Body: PKCS#7 SignedData, version 1, digest algorithm SHA‑256, content type 1.3.6.1.4.1.311.2.1.4 (SPC_INDIRECT_DATA_OBJID, i.e. an Authenticode indirect-data descriptor), signature algorithm 1.2.840.10045.2.1 (ECDSA).

3.7.2 Path

[0] SSL.com Root Certification Authority ECC           self-signed root
     └─ SSL.com Code Signing Intermediate CA ECC R2     CA=true, 2019-03-07 → 2034-03-03
          └─ AN Soft                                    leaf,   2026-06-11 → 2027-06-11
[1] DigiCert Trusted Root G4                            (timestamping chain, separate)
     └─ DigiCert Trusted G4 TimeStamping RSA4096 SHA256 2025 CA1   CA=true
          └─ DigiCert SHA256 RSA4096 Timestamp Responder 2026 1    leaf, TSA

3.7.3 Leaf certificate

Subject        CN=AN Soft
               O=AN Soft
               L=Bergen op Zoom
               ST=North Brabant
               C=NL
Issuer         CN=SSL.com Code Signing Intermediate CA ECC R2
               O=SSL Corp, L=Houston, ST=Texas, C=US
Serial         158467513303841965040394699300328354453
NotBefore      2026-06-11 15:40:51 UTC
NotAfter       2027-06-11 15:40:51 UTC
Key usage      digitalSignature
Extended KU    code_signing (1.3.6.1.5.5.7.3.3)
Public key     ECC
Signature      sha256-ecdsa

Characteristics of note:

  • OV, not EV. The subject is organisation-validated (the legal name is checked) but there is no jurisdictionOfIncorporation* attribute set, no EV policy OID, and the validity is exactly one year. EV certificates are normally issued for two to three years and carry jurisdictional attributes. A one-year OV is the cheapest identity a company can present to SmartScreen.
  • ECC leaf. Modern, small, fast to verify, and — operationally — not the RSA key generation that featured in the 2022–2023 public-CA key-strength deprecations.
  • No serialNumber / organizationIdentifier attribute in the subject. For a Dutch eenmanszaak or vof this is normal (no statutory registration number is required in the DN), but it means the certificate does not itself disclose the KvK number that the vendor publishes on its website (§8.1). The certificate and the company register must be joined by the analyst, not by the validator.

3.7.4 Counter-signature

An RFC‑3161 signature_time_stamp_token is present as an unsigned attribute over the SHA‑256 digest of the signature value:

TSA leaf      CN=DigiCert SHA256 RSA4096 Timestamp Responder 2026 1
              O=DigiCert, Inc., C=US
              serial 11048481613454637687985428644217624629
              valid  2026-08-05 → 2037-11-04
TSA issuer    CN=DigiCert Trusted G4 TimeStamping RSA4096 SHA256 2025 CA1
              serial 18316726489210163671024321415498844550
Root          DigiCert Trusted Root G4  (issuer: DigiCert Assured ID Root CA, 2031)

The counter-signature is what makes the Subject verifiable after 2027‑06‑11, when the leaf expires. Windows will continue to report "digitally signed by AN Soft, timestamped" indefinitely, even if the certificate has been revoked for future signing and even if the vendor has ceased trading.

3.8 Component composition

The Subject is 96.2 % library and 3.8 % application code by volume.

Layer Approx. share Identification
UCRT + C++ runtime + EH + locale/iostream ~68 % MSVC static CRT; three ASCII literal pools and three UTF‑16 pools of locale data
std::filesystem / std::error_code implementation ~9 % full system-error string table ("cross device link", "owner dead", …)
PE / debug / RTTI machinery ~5 % .pdata + .xdata, type descriptors, COMDAT tables
GDI+ / WinINet import thunks ~4 % 12 + 10 functions, plus their initialisation
Application code ~3.8 % 28 KB across ~35 functions in 0x140003000–0x140035000

Decompilation recovered 1 654 functions; approximately 35 belong to the application. All application-authored human-readable text in the file amounts to under 5 KB (§Appendix B).

3.9 Library fingerprint

Component Evidence Role
nlohmann/json 3.11.2 banner 961c151d2e87f2686a955a9be24d316f1362bf21 3.11.2; [json.exception.; , column ; at line; type must be string, but is primary feed parser (JSON)
rapidxml .?AVparse_error@rapidxml@@; expected <; expected element name; expected >; expected attribute name; expected =; expected ' or "; expected ;; invalid numeric character entity; unexpected end of data; bad conversion alternative feed parser (XML), same schema
WinReg (Dicanio header library) .?AVRegException@winreg@@; ADVAPI32 imports are exactly RegCreateKeyExW, RegSetValueExW, RegGetValueW, RegCloseKey registry RAII wrapper
GDI+ GdiplusStartup, GdipLoadImageFromStream, GdipCreateFromHDC, GdipDrawImageRectI, GdipCloneImage, .?AVImage@Gdiplus@@ streamed image rendering
WinINet 10 imports incl. InternetCrackUrlA, HttpOpenRequestA, HttpSendRequestA all HTTP
urlmon URLDownloadToFileW, URLOpenBlockingStreamA, .?AVDownloadCallback@@, .?AUIBindStatusCallback@@ file download + progress
comctl32 v6 InitCommonControlsEx, msctls_progress32, SysLink progress bar, hyperlinks
Shell32 ShellExecuteW, ShellExecuteExW, SHGetKnownFolderPath open file / open folder

The presence of two complete feed parsers, one JSON and one XML, with the same field vocabulary indicates the backend can serve either representation; &comma; in the literal pool (0xA05B0) is the XML entity the server uses to carry a literal comma inside an attribute or text node, i.e. XML is a first-class serialisation and not a vestigial include.

3.10 Capability surface (imports)

Ten DLLs, 187 functions. ADVAPI32.dll contributes four — all registry. The complete import table is reproduced in Appendix A. The distribution of functions across libraries is itself informative:

DLL fns Reading
KERNEL32 103 of which ~80 are UCRT support (Fls*, RtlVirtualUnwind, GetStringTypeW, heap, locale, console, file-buffer); ~23 are application-visible (CreateProcessW, CreateThread, GetTempPathW, GetTempFileNameW, CreateDirectoryW, DeleteFileW, GetFileAttributesW, WaitForSingleObject, GetExitCodeProcess, Sleep, …)
USER32 36 one window class, custom-drawn controls, TrackMouseEvent, SetWindowLongPtrW
GDI32 13 GDI painting only (fonts, pens, brushes, RoundRect); no bitmaps, no DIBs
ADVAPI32 4 registry only
SHELL32 3 ShellExecuteW, ShellExecuteExW, SHGetKnownFolderPath
ole32 1 CoTaskMemFree (release of SHGetKnownFolderPath memory)
urlmon 2 URLDownloadToFileW, URLOpenBlockingStreamA
gdiplus 12 image decode/scale/draw
WININET 10 HTTP
COMCTL32 3 InitCommonControlsEx + 2 ordinals

Capabilities the Subject does not link, and therefore cannot perform without run-time symbol resolution (none of which is present — no LoadLibrary/GetProcAddress pair is imported by the application layer, only by the UCRT's own machinery):

Sockets / TCP / UDP / DNS .............. ws2_32  NOT IMPORTED        -> no network path except WinINet/urlmon
Credential-store cryptography .......... no CryptUnprotectData      -> crypt32 NOT IMPORTED
System crypto primitives ............... bcrypt  NOT IMPORTED       -> no hashing in-process
Named pipes / RPC ...................... kernel32 CreateNamedPipeW not imported
Services ............................... advapi32 CreateServiceW not imported
WMI .................................... no COM class instantiation (ole32 = CoTaskMemFree only)
Token / privilege / SID inspection ..... advapi32 OpenProcessToken/GetTokenInformation NOT imported
Process / module enumeration .......... kernel32 CreateToolhelp32Snapshot NOT imported
Thread suspend / injection ............. kernel32 SuspendThread, OpenThread NOT imported
Registry enumeration (Run keys, etc.) .. advapi32 RegEnumKeyEx/RegEnumValueW NOT imported
Shell / startup-folder writes ........... none

The registry wrapper's fixed four-function import set is decisive: because WinReg is the only registry interface the program has, and WinReg's API surface is Create/SetValue/GetValue/Close on keys named by the caller, the Subject is structurally incapable of enumerating the registry or of discovering an autorun entry it was not told about. It reads a fixed, short list of keys, by name, as part of the targeting model of §4.4 — and writes exactly two.


4. Architecture and control flow

4.1 Application layer

Of 1 654 recovered functions, ~35 implement the program. The remainder are UCRT, the C++ standard library, locale facets, the exception/unwind machinery, GDI+ and WinINet thunks, and the std::filesystem/std::error_code implementation.

RVA Size Recovered role
0x14000C990 971 wWinMain
0x14000D2E0 — WndProc (owner-drawn non-client, hit-testing, hover state, paint)
0x140003E00 678 Bootstrap — descriptor fetch, fallback, config application, wizard construction
0x14000A700 8 069 ApplyConfig(json) — parse, targeting evaluation, page vector construction
0x140004880 6 928 ResolveProduct — GetModuleFileNameW, stem/token handling, identifier extraction
0x140008180 1 038 HttpPost(url, body) — WinINet request/response
0x140007D70 1 036 HttpGet(url) — InternetOpenUrlA variant, 2 000-byte reads
0x14002EAC0 468 HttpGet2(url) — second GET path (geo, images)
0x14002E3B0 731 TapAction(key) — funnel beacon builder
0x140021910 368 BuildGetDownloadInfo()
0x1400040B0 873 OpenOfferLink() — compose deep link, ShellExecuteW, return for storage
0x140007040 243 URLDownloadToFileW wrapper
0x14000CD70 430 LoadImageFromUrl() — URLOpenBlockingStreamA → GdipLoadImageFromStream
0x140008C00 1 301 RunSilent_ADR() — brand-specific download/execute
0x140009120 1 301 RunSilent_GLY() — byte-identical twin
0x1400304E0 814 EnsureDirectory()
0x140030810 527 ConsentAccept()
0x140030A20 1 396 OfferPageNext() — accept/decline, receipt, survey branch
0x140031440 1 026 MainWndProc
0x140030FB0 1 160 BuildWizardPage()
0x14000FE30 1 092 page teardown loop
0x14000E1E0 771 BuildEulaPage()
0x14000E4F0 886 BuildTextPage()
0x14000E870 1 001 BuildPrivacyPage()
0x1400039D0 567 RegGetString() (winreg wrapper)
0x14000A470 646 SpawnProc() — CreateThread + ShellExecuteW
0x140006C50 1 002 OpenFolder() — ShellExecuteExW, explorer /select,
0x14000EC70 — to_wstring(DWORD) — error formatting for the single failure dialog

Two functions — RunSilent_ADR and RunSilent_GLY — are 1 301 bytes each and differ only in the three-character temp prefix and the registry brand string. They are the same source function instantiated twice because those strings are compile-time literals rather than parameters. This is the first evidence of the dual-brand structure developed in §4.8.

4.2 Entry sequence

wWinMain(hInst, hPrev, cmdline, show)
{
    WNDCLASSEXW wc = {};
    wc.cbSize = 0x50;
    wc.style = CS_HREDRAW | CS_VREDRAW;               /* 3 */
    wc.lpfnWndProc = WndProc;
    wc.hInstance = hInst;
    wc.hCursor = LoadCursorW(NULL, IDC_ARROW);       /* 0x7F00 */
    wc.hbrBackground = (HBRUSH)(COLOR_WINDOW + 1);    /* 6 */
    wc.lpszClassName = L"Installer";                 /* 0xA09C0 */

    if (!RegisterClassExW(&wc)) {
        std::wstring msg = L"Window Class Registration Failed! Error code: ";
        msg += std::to_wstring(GetLastError());
        MessageBoxW(NULL, msg.c_str(), L"Error", MB_ICONERROR);
        return 0;
    }

    ApplyConfig(&g_defaultConfig);                    /* (1) static defaults */
    if (!g_gdiplusInit) {                             /* (2) lazy GDI+      */
        GdiplusStartupInput in(1, NULL, FALSE);
        GdiplusStartup(&g_gdiToken, &in, NULL);
        g_gdiplusInit = true;
    }
    Bootstrap();                                      /* (3) NETWORK       */

    int cx = GetSystemMetrics(0), cy = GetSystemMetrics(1);
    g_fontTitle = CreateFontW(24,0,0,0,600,0,0,0,DEFAULT_CHARSET,0,0,0,0x20,L"Calibri");
    g_fontH1    = CreateFontW(18,0,0,0,600,0,0,0,DEFAULT_CHARSET,0,0,0,0x20,L"Segoe UI");
    g_fontBody  = CreateFontW(16,0,0,0,400,0,0,0,DEFAULT_CHARSET,0,0,0,0x20,L"Segoe UI");
    g_fontSmall = CreateFontW(12,0,0,0,400,0,0,0,DEFAULT_CHARSET,0,0,0,0x20,L"Segoe UI");

    g_hWnd = CreateWindowExW(0, L"Installer", L"Installer",
                             WS_POPUP | WS_VISIBLE,          /* 0x90000000 */
                             (cx - 692) / 2, (cy - 492) / 2,
                             692, 492, NULL, NULL, hInst, NULL);

    MSG msg;
    while (GetMessageW(&msg, NULL, 0, 0)) {
        TranslateMessage(&msg);
        DispatchMessageW(&msg);
    }
    return 0;
}

Observations:

  1. Defaults are applied before anything is known. ApplyConfig is called with a statically initialised nlohmann::json at 0x1400B6E00, and then again from Bootstrap() with the server payload. The two calls take the same code path with the same type, so the "hard-coded fallback" is a genuine configuration document, not a scattering of defensive ifs. It is reachable for offline inspection (see OQ‑4).
  2. GDI+ is initialised before the network call because the interface must be able to render the first page regardless of connectivity. The network call is blocking and on the UI thread (3).
  3. Bootstrap() is synchronous on the UI thread and runs to completion before the window is created. Consequence: no window exists while DNS, TCP, TLS and the descriptor exchange are outstanding. On a slow or blackholed path the process presents as a running process with no window — indistinguishable from a hang, and a common source of "nothing happened" reports. There is no timeout constant in the literal pool, so the effective timeout is the WinINet default (30 s connect / 30 s send / 30 s receive per request, doubled by retries in some paths).
  4. Typography is fixed: Calibri 24 semibold, Segoe UI 18 semibold, Segoe UI 16 regular, Segoe UI 12 regular. FW_SEMIBOLD = 600. There is no font fallback, no scaling by DPI and no user preference; on a system without Segoe UI (allowed, but unusual) text renders in the generic face at these pixel sizes.
  5. GetMessageW return value is not validated for the -1 error case, and wParam/lParam are discarded in TranslateMessage. Neither is consequential.
  6. WS_POPUP | WS_VISIBLE ⇒ no caption, no border, no system menu, no minimise/maximise box, no WM_SIZE handling. The user cannot move the window, cannot resize it, cannot close it with the title-bar X (there is none) and — because hIcon is NULL — it has no taskbar icon distinct from the default. Termination is via the custom close control, Alt+F4 (which arrives as WM_CLOSE), or Task Manager.

4.3 Interface model

The window is a 692 × 492 owner-drawn card, centred on the primary monitor:

Aspect Implementation Evidence
Frame borderless WS_POPUP, COLOR_WINDOW+1 brush §4.2, class registration
Child controls created per page and destroyed on page change GetWindow(GW_CHILD) / GW_HWNDNEXT loop in OfferPageNext
Button rendering RoundRect + CreatePen + CreateSolidBrush + SetBkMode(TRANSPARENT) + DrawTextW GDI32 import set (13 functions, no bitmap API)
Hover state TrackMouseEvent + InvalidateRect 0x14000D220
Text DrawTextW with SetTextColor/SetBkColor USER32/GDI32 imports
Progress msctls_progress32 (PBM_SETPOS/PBM_SETRANGE via SendMessageW) comctl32 import + class literal
Hyperlinks SysLink class via InitCommonControlsEx(ICC_LINK_CLASS) SysLink literal + control creation
Inputs STATIC, BUTTON, EDIT, COMBOBOX control class literals
Survey host private SurveyWinClass / SurveyHost class registration literals

The interface is fully owner-drawn: there are no dialog templates (§3.3), no stock controls on the shell, and no visual identity. Text is set through SetWindowTextW from wide strings that originate in the feed (§6). The practical consequence is that the Subject's appearance is a function of the server response, not of the binary: titles, body copy, legal links, offer names and even the number and order of pages are data.

The 692 × 492 fixed metric with no AdjustWindowRect and no DPI rescaling (despite the True/PM manifest declaration, §3.4) means the interface is not resolution- or scaling-independent: at 200 % scaling it occupies a quarter-screen card with 16-pixel body text; at 1280 × 720 it is nearly full-height.

4.4 Configuration model and targeting engine

ApplyConfig evaluates, for each element of the feed's offers array, three predicates. Offers surviving all three become pages of the wizard.

void ApplyConfig(const json& j)
{
    g_channel = get_string(j, "channel");
    g_user_id = get_string(j, "userID");

    for (const json& o : j["offers"]) {
        /* P1 - geography */
        if (o.contains("targetGeos")) {
            bool ok = false;
            for (auto& g : o["targetGeos"])
                if (iequals(g.get<std::string>(), g_country)) ok = true;
            if (!ok) continue;
        }
        /* P2 - registry presence */
        if (o.contains("registryChecks")) {
            bool present = true;
            for (auto& c : o["registryChecks"]) {
                if (!RegKeyValueExists(c["root"], c["path"], c["value"]))
                    present = false;
            }
            if (present) continue;
        }
        /* P3 - filesystem presence */
        if (o.contains("fileChecks")) {
            bool present = true;
            for (auto& f : o["fileChecks"]) {
                std::wstring p = ExpandEnvironmentStrings(widen(f.get<std::string>()));
                if (GetFileAttributesW(p.c_str()) == INVALID_FILE_ATTRIBUTES)
                    present = false;
            }
            if (present) continue;
        }
        g_pages.push_back(Page::FromJson(o));
    }
}

Semantics recovered from the call sites and the served data:

  • targetGeos is an allow-list of ISO‑3166‑1 alpha‑2 codes. Its negation form is a single leading ! (["!DE"] = "everywhere except Germany"), which is the form used by two of the five served offers (§7.5). An empty array ([]) is treated as "no restriction". Only one country is compared, so the model is single-geo, not "geo-set intersection": a user behind a VPN is offered the bundle for the VPN's exit country.
  • registryChecks is a suppression list: if every {root,path,value} triple exists, the offer is hidden. It is used exclusively to avoid re-offering software the user already has.
  • fileChecks performs the same role in the filesystem, with %ProgramFiles% expansion and, in the served data, a bare path with no executable name — the check is for the directory.
  • All three predicates are monotone towards hiding: nothing in the engine promotes an offer. A feed that returns fewer offers than expected can therefore be caused by any of the three, and the client has no telemetry path that reports which.

iequals (a case-insensitive wide-string compare) is the comparison used for country codes; the implication is that the server's targetGeos values are matched without canonicalisation, so de, DE and De are all valid and all match.

4.5 Elevation model

The manifest (§3.4) requests requireAdministrator unconditionally. The offers do not require it — and this is demonstrable from the server's own arguments, not from inference:

Offer Installer arguments Scope implied
OperaGX --silent --allusers=0 explicitly per-user
AIChat /S NSIS silent default; per-user unless /ALLUSERS is passed (it is not)
Glory --silent --launch no scope switch
AdReject --silent --channel IU no scope switch
Premier Opinion -c:1538 -t:InstallUnion no scope switch

--allusers=0 is the strongest single datum: the only offer whose installer documents an install scope is configured not to use the machine scope. A per-user installation succeeds without elevation. Therefore:

Finding F‑20 — Elevation is acquired speculatively. The Subject takes Administrator before it has contacted the network, before it knows which offers apply, and before any of those offers has expressed a need for it. Four of the five offers can be delivered per-user. The token's purpose is not to satisfy a known requirement; it is to remove the requirement from the design space of any future offer the server may return — including offers that did not exist when this build was signed. The user's sole consent gate (the UAC dialog) is presented before the offer is named, and is never re-presented as the offer set changes. Severity: Moderate‑High. Confidence: High (manifest, five argument vectors, and the absence of any IsUserAnAdmin-style runtime check).

This has an architectural corollary which is developed in §10: the highest-privilege decision in the process is made with the least information. Everything after WinMain is data-driven, and all of it inherits the token that was granted to a window class registration.

4.6 Download and execution engine

RunSilent_ADR / RunSilent_GLY are the only functions that write an executable to disk and start it.

void RunSilent_<BRAND>(const std::wstring& url, const std::wstring& extra_args)
{
    wchar_t tempDir[MAX_PATH], tmpName[MAX_PATH];

    /* 1. acquire a UNIQUE NAME, not a temporary file */
    GetTempPathW(MAX_PATH, tempDir);                       /* %TEMP%\              */
    GetTempFileNameW(tempDir, L"ADR", 0, tmpName);         /* %TEMP%\ADRx123.tmp   */
    DeleteFileW(tmpName);                                 /* delete the .tmp      */
    std::wstring exe = std::wstring(tmpName) + L".exe";    /* ADRx123.tmp.exe      */

    /* 2. fetch */
    DeleteFileW(exe.c_str());
    HRESULT hr = URLDownloadToFileW(NULL, url.c_str(), exe.c_str(), 0, NULL);
    if (FAILED(hr)) { DeleteFileW(exe.c_str()); return; } /* silent, no retry */

    /* 3. compose */
    std::wstring cmd = L"\"" + exe + L"\" --silent";
    if (!extra_args.empty()) cmd += L" " + extra_args;    /* unquoted */

    /* 4. record consent BEFORE the child runs */
    HKEY hk; DWORD one = 1;
    RegCreateKeyExW(HKEY_CURRENT_USER, L"SOFTWARE\\<Brand>", 0, NULL, 0,
                    KEY_WRITE /*0x20006*/, NULL, &hk, NULL);
    RegSetValueExW(hk, L"Consent", 0, REG_DWORD, &one, 4);
    RegCloseKey(hk);

    /* 5. run synchronously, windowless */
    STARTUPINFOW si = {0}; si.cb = sizeof(si);
    PROCESS_INFORMATION pi = {0};
    if (CreateProcessW(exe.c_str(), cmd.data(), NULL, NULL, FALSE,
                       CREATE_NO_WINDOW /*0x08000000*/, NULL,
                       tempDir /* cwd */, &si, &pi)) {
        WaitForSingleObject(pi.hProcess, INFINITE);
        GetExitCodeProcess(pi.hProcess, &code);
        CloseHandle(pi.hThread); CloseHandle(pi.hProcess);
    }
}

Properties of this routine that are relevant to assessment:

  1. The .tmp file is created and immediately deleted. GetTempFileNameW is used as a uniqueness oracle, not as a file creator. The resulting path contains .tmp in its middle and .exe at its end, so the extension is .exe (loadable/executable) while naive pruning rules that match *.tmp are not triggered by the name the user sees. Registry-free, path-obfuscated staging.
  2. No content verification. URLDownloadToFileW reports transport success. There is no length check against a declared value, no hash, no WinVerifyTrust, no MapAndLoad+signature, no PE header test, no version check. A transport-200 response of 0 bytes, of an HTML error page, or of a later-dated binary is equally acceptable. If the result is not a valid image, CreateProcessW fails and the failure is discarded.
  3. No quoting of the appended arguments. The path is quoted ("…exe"); --silent and any params from the feed are concatenated unquoted. CreateProcessW reparses the whole string with the familiar CommandLineToArgvW rules, so spaces inside params become token boundaries, and a " inside params opens a quoted section. This is an intentional interface — the server writes a command line, the client assembles it — but the interface is string-level, with no escaping or structured encoding (§10.4).
  4. Consent is written before the child is launched, and on the download-success path only. A user who downloads the offer and then kills it at the first dialog, or whose antivirus quarantines the child before its first instruction, is recorded as having consented. The value is a click receipt, not an install confirmation.
  5. CREATE_NO_WINDOW, not DETACHED_PROCESS. The child has a console; it is hidden. Children of the child inherit the hidden console unless they allocate their own, which means console-oriented installers run without the customary flash but also without the customary indication that anything is happening.
  6. WaitForSingleObject(INFINITE) and the exit code read mean the parent can distinguish 0/3010-style success and reboot-required codes, and does so (the code is stored to a global used by the Finish page). The operation is synchronous by design.

The child's lifetime is bounded by the parent's; closing the Subject's window with the close control while a child is running terminates the owner-drawn close path but not the process tree, so the child continues. This is observable as "an installer window closed but something kept installing".

4.7 Telemetry model

Two independent collectors.

(a) Ad-network collector — post.securestudies.com/TapAction.aspx (TAP), over plain HTTP:

http://post.securestudies.com/TapAction.aspx?
    [&uid=<16-hex installation id>]
    campaign_id=1538
    &tpi=InstallUnion
    &action_id=<N>
    [&id=<offer id>]          (action_id=8 only)
    [&consent_version=<v>]   (action_id=1 only)
action_id Emitted at Recovered meaning
0 first page construction launch
1 legal/privacy page shown consent presentation (+ version)
2 page transition page view
3 catalogue page rendered offers shown
4 offer page rendered offer impression
8 Accept on the optional offer conversion (+ offer id)
9 Decline, or single-page path opt-out
17 finish page completion

Zero-width GIF responses are discarded (HttpGet does not inspect content type); the beacon is fire-and-forget and the interface does not depend on it succeeding. Killing securestudies.com changes nothing observable except the vendor's statistics.

(b) Vendor collector — dlsft.com/callback.php and the dpdv3 pixel (§5.1). Served over HTTPS, so it survives DNS filtering that kills (a). callback.php carries act=started|completed and the channel/id pair from the product descriptor; dpdv3 carries campaignid=1538 and a co= value which the code supplies as <Blocked> when the corresponding datum is unavailable — the string <Blocked> therefore appears in the vendor's own access logs as a literal parameter value, which makes the two codes distinguishable in the log without any other instrumentation.

The dual-collector design is standard for the class and is not evidence of concealment: the ad network is paid per event and requires its own pixel; the network operator is paid per download and requires its own. An analyst who blocks one and not the other sees a "partially silent" client, which is worth knowing before drawing conclusions from absence of traffic.

4.8 State, persistence and the dual-brand structure

Persistent state touched by the application layer:

Store Name Type Written when Read when
HKCU SOFTWARE\AdReject\Consent REG_DWORD = 1 offer 2 accepted, before child launch nothing (vendor-side check)
HKCU SOFTWARE\Glory\Consent REG_DWORD = 1 offer 3 accepted, before child launch nothing (vendor-side check)
HKCU SOFTWARE\InstallUnion\1538\BundleOfferActionUid REG_SZ offer page completed (≥ 2 pages) next launch, to suppress
%TEMP% ADR<hex4>.exe, GLY<hex4>.exe PE download execute, then delete
%LOCALAPPDATA% adreject\file.txt, glory\file.txt data settings/state write read

There is no autostart: no Run/RunOnce, no startup-folder shortcut, no scheduled task, no service. The registry footprint is write-once-per-install, and deleting it does not break the application (the values are suppression flags; their absence merely means the offer may be presented again).

The two brand strings are the only indication in the binary that this codebase serves more than one product. They are not user-visible; they are the same product published under two names, with disjoint consent keys so that accepting one does not suppress the other:

brand A    AdReject      %TEMP% prefix "ADR"    HKCU\SOFTWARE\AdReject    %ProgramFiles%\AdReject
brand B    Glory        %TEMP% prefix "GLY"    HKCU\SOFTWARE\Glory       %ProgramFiles%\Glory
shared     InstallUnion\1538                  BundleOfferActionUid

This is confirmed independently in §8: both AdReject and Glory are published, signed, versioned and priced by the same organisation, which is also the organisation that built the Subject.

4.9 Failure modes

Condition Observed/derived behaviour
No network, DNS blackholed Bootstrap blocks to WinINet timeout; falls back to DEFAULT_CONFIG; wizard renders with default page set; no offer payloads resolvable
HTTP 404/500 on service.php HttpPost returns body; parse fails; fallback config
Malformed JSON nlohmann throws parse_error; fallback path; no dialog
offers missing or null wizard has zero offer pages; decline event action_id=9 fires
payload URL unreachable URLDownloadToFileW fails; temp file deleted; no user-visible error
child exits non-zero exit code recorded; Finish page displays regardless
user clicks close during child run window destroyed; child continues (process-tree ordering, §4.6)
GetTempFileNameW fails (MAX_PATH/permission) goto cleanup; no dialog, silent return

The Subject's error strategy is uniform: fail silently, render something, never block the funnel. The single exception is the RegisterClassExW failure dialog (§4.2), which is the only error path a user can see. Errors are otherwise converted into short-lived wide strings in a helper (0x14000FA20 prefix-joins "<message> failed: ", 0x1400128B0 appends the numeric GetLastError) and dropped.


5. Network infrastructure

5.1 Endpoint inventory

The Subject contains five absolute URLs (rendered in Appendix B with RVAs). Combined with the URLs in the payloads it retrieves and in the assets it references, the observed infrastructure is:

# URL / host Method Purpose Transport Resolver
1 https://www.dlsft.com/geo/ GET bootstrap: issue installation identifier + resolve country TLS 1.2+ Cloudflare
2 https://dlsft.com/service.php?request=offers POST offer bundle for the installation TLS Cloudflare
3 https://www.dlsft.com/service.php GET/POST alternate base used by the deep-link composer TLS Cloudflare
4 https://dlsft.com/service.php?request=getDownloadInfo&name=<name> POST product descriptor TLS Cloudflare
5 https://dlsft.com/callback.php GET vendor-side event receipt (`act=started completed`) TLS
6 https://dlsft.com/PO.jpg GET promotional creative TLS Cloudflare
7 https://filedm.com/privacy.php GET Privacy Policy (hyperlinked in interface) TLS Cloudflare
8 https://filedm.com/terms.php GET Terms of Service / EULA (hyperlinked) TLS Cloudflare
9 http://post.securestudies.com/TapAction.aspx GET funnel telemetry cleartext HTTP Cloudflare
10 https://dpd.securestudies.com/dpdv3.aspx GET post-download pixel TLS Cloudflare
11 http://directfiledl.com/getfile GET product archive cleartext HTTP, no 443 listener Hetzner

5.2 Request construction

The application speaks WinINet directly; there is no HTTP library, no cookie jar, no connection pooling, no proxy configuration and no user-agent rotation.

/* outbound */
InternetOpenA("Mozilla/5.0", INTERNET_OPEN_TYPE_DIRECT /* 1 */, NULL, NULL, 0);

/* offers + getDownloadInfo are POSTs with a form body */
InternetCrackUrlA(url, 0, 0, &components);          /* ICU flags 0x68 */
InternetConnectA(session, components.lpszHostName,
                 (scheme == 4) ? 443 : 80,           /* HTTPS -> 443        */
                 NULL, NULL, INTERNET_SERVICE_HTTP /* 3 */, 0, 0);
HttpOpenRequestA(conn, L"POST", path, NULL, NULL, NULL,
                 secure ? 0xA4C00000 : 0xA4400000);  /* SECURE|RELOAD|…     */
HttpSendRequestA(req, headers, header_len, body, body_len);
   headers = "Content-Type: application/x-www-form-urlencoded"
for (;;) { InternetReadFile(req, buf, 0x800 /* 2048 */, &n); if (!n) break; append(buf, n); }

Points of analytic substance:

  • INTERNET_OPEN_TYPE_DIRECT (1) — not PRECONFIG (0). The Subject therefore ignores system and per-user proxy configuration, WinHTTP autoproxy, PAC files and WinINet LAN settings. An environment that has enforced egress filtering by requiring a proxy will see this client either fail or attempt an unauthorised direct path. For the analyst it is a useful discriminator: a proxy-based capture that shows no connection is not proof of no connection.
  • No cookie persistence. HttpOpenRequestA is called with a NULL state array and without INTERNET_FLAG_NO_COOKIES, so WinINet may create an in-memory/session cookie jar for the process, but nothing is written to %TEMP% and nothing survives exit. Identity is carried in URLs, not cookies.
  • Mozilla/5.0 as both User-Agent and session name. The UA has no platform token, no AppleWebKit, no (KHTML, like Gecko), no version. It is a 10-character constant. Against any real browser population this is a low-entropy, high-precision signature (§11.2).
  • Response cap. InternetReadFile into a 2 048-byte stack buffer, appended to a std::string, with no maximum size. A malicious or broken server could drive an unbounded heap growth; there is no guard. In practice the feed is 8.9 KB.
  • Error strings are built but never surfaced ("InternetOpenA failed: ", "Failed to parse URL: ", "InternetConnectA failed: ", "HttpOpenRequestA failed: ", "HttpSendRequestA failed: ", "Error reading from server: "). They are consumed by a helper that appends GetLastError() and discards the result. There is no log file, no OutputDebugString call site, and no dialog. Debugging the Subject's network behaviour in situ therefore requires a debugger or an API monitor; the application will not tell you what went wrong (§4.9).

5.3 Live interrogation

All ten endpoints were resolved and interrogated on 2026‑11‑16/17 from a United Kingdom egress, with User-Agent: Mozilla/5.0.

Endpoint Status Bytes Content-Type Body as served
GET https://www.dlsft.com/geo/ 200 18 text/html; charset=UTF-8 3482d239f6935cd,GB
GET https://dlsft.com/ 200 2 text/html US
GET https://dlsft.com/service.php?request=offers&channel=default 200 0 text/html (empty)
POST …request=offers 200 8 887 text/html JSON, 5 offers
POST …request=getDownloadInfo&name=ByteBreaker V1_85892839 200 232 text/html JSON, 1 product
GET https://dlsft.com/callback.php 200 0 text/html (empty)
GET https://dlsft.com/PO.jpg 200 32 755 image/jpeg 413 × 30 progressive JPEG
GET https://filedm.com/privacy.php 200 9 235 text/html rendered policy
GET https://filedm.com/terms.php 200 7 635 text/html rendered terms
POST http://directfiledl.com/getfile 200 12 text/html not valid id

Two operational details are confirmed by the table and matter in practice:

  • service.php is POST-only. A GET with the same query string returns 200 and a zero-length body. Any analyst who reproduces the protocol with GETs will conclude, wrongly, that the service is dead. (§7.4 of the companion notes made precisely this inference and had to be corrected.)
  • directfiledl.com has no TLS listener. curl https://directfiledl.com/getfile fails at connect() after ~1.2 s; curl http://… succeeds. The service is not merely configured to serve plain HTTP as a default — HTTPS is unavailable, which means a client that upgrades (HSTS, Upgrade-Insecure-Requests, a .app domain) cannot fetch the product at all.

/geo/ returns a CSV pair. The first field is a 16-character lowercase hexadecimal installation identifier; the second is an ISO‑3166‑1 alpha‑2 country code computed by the server from the client's IP address. The identifier is 64-bit, not a UUID (a v4 UUID would be 32 hex with hyphens), not a hash of anything client-side (it is stable only for the request in which it is issued; see §5.5), and it is the only user-scoped token in the entire protocol.

5.4 The "security code" scheme

directfiledl.com/getfile is not a static file path. It validates two parameters, and the second is computed by the client from the first. Reconstruction of the generator (0x140009CF0, 1 022 B) and of its caller (0x14000A470, 646 B):

std::wstring SecurityCode()
{
    static const wchar_t* SALT = L"SCI_multipart";      /* RVA 0xA08A0, 13 chars */
    std::wstring g = g_id;                              /* JSON "id", wide      */
    std::wstring x;
    for (size_t i = 0; i < g.size(); ++i)
        x.push_back((wchar_t)(g[i] ^ SALT[i % 13]));     /* repeating-key XOR    */

    uint32_t table[256];
    for (uint32_t n = 0; n < 256; ++n) {                 /* CRC-32, reflected    */
        uint32_t c = n;                                 /* poly 0xEDB88320      */
        for (int k = 0; k < 8; ++k) c = (c & 1) ? (c >> 1) ^ 0xEDB88320 : c >> 1;
        table[n] = c;
    }
    uint32_t crc = 0xFFFFFFFF;
    for (wchar_t ch : x) crc = (crc >> 8) ^ table[(ch ^ crc) & 0xFF];   /* low byte per char */
    crc = ~crc;

    std::wstringstream ss;
    ss << std::hex << std::setfill(L'0') << std::setw(8) << crc;
    return ss.str();
}

/* caller */
url = g_payload + L"?id=" + g_id + L"&s=" + SecurityCode();

Worked example, g_id = L"85892839":

id        38 35 38 39 32 38 33 39                        "85892839"
salt      53 43 49 5F 6D 75 6C 74                        "SCI_mult"
xor       6B 76 71 66 5F 4D 5F 4D                        "kvqf_M_M"
crc32     6F 75 A7 CD
s         6f75a7cd
url       http://directfiledl.com/getfile?id=85892839&s=6f75a7cd
$ curl -sS -D- -o bytebreakerV1.zip \
    'http://directfiledl.com/getfile?id=85892839&s=6f75a7cd'
HTTP/1.1 200 OK
Content-Length: 16787263
Content-Type: application/octet-stream
X-Powered-By: Express

$ sha256sum bytebreakerV1.zip
aa3fc2c089a8f039b2198d6f64e4e727e1631c7e18fe084429b1b491330ad85f  bytebreakerV1.zip

Negative cases:

GET /getfile                                    HTTP/1.1 200  →  "not valid id"                                   (12 B)
GET /getfile?id=9999                            HTTP/1.1 200  →  "not valid id"
GET /getfile?id=abc                            HTTP/1.1 200  →  "not valid id"
GET /getfile?id=85892839                        HTTP/1.1 200  →  "Security code is not matching: carefully copy and do not change the link." (73 B)
GET /getfile?id=85892839&s=6f75a7cd             HTTP/1.1 200  →  ZIP, 16,787,263 B, application/octet-stream

5.4.1 Assessment of the scheme

Property Value Consequence
Key fixed, compiled into every copy recoverable by anyone with a copy (this report)
Cipher repeating-key XOR over a 13-byte salt not a cipher; a mask
Tag CRC‑32 (linear, 32-bit, unkeyed-combinable) not a MAC; forgery is trivial
Input the public id only no per-user, per-build or per-time component
Output length 8 hex characters, zero-padded 32 bits
Effect prevents other sites embedding getfile?id= ✓
Effect prevents URL scanners/unfurlers fetching payloads ✓
Effect prevents an analyst reading a packet capture ✗
Effect binds a URL to a recipient or a time ✗
Effect prevents hotlink enumeration ✗ (one valid s per id, computable once the law is known)

The scheme is best characterised as link signing as DRM, not as access control. Its economics are sensible: an unreasoned crawler, a Discord embedder or a Lean/Threads "link unfurl" bot will not compute the code, so 99 % of the accidental traffic that would otherwise consume bandwidth is deflected. Its weakness is that the entire security rests on a constant inside a file the user already has. There is no rate limit, no short-circuit cache, no per-IP binding and no nonce, all of which are cheap to add.

5.4.2 The validator is a shape test, not a record test

?id= was probed with nine values to distinguish "identifier exists" from "identifier is well formed":

id= Response
85892839, 8589283, 85892840 Security code error ⇒ passes
9999, 0, 1, abc, `` (empty) not valid id ⇒ fails

Both 7-digit and 8-digit values pass, and consecutive values around the subject pass, although the subject is the only identifier in a 701-value window that resolves to a product (§6.2). The check is therefore a pattern test (\d{7,8}, or a length comparison) executed before the signature check, and it discloses two properties of the backend that were not intended to be disclosed: the key is a short integer, and the signature check runs before the database lookup (so ?id=…&s=… for an existent product costs one row read, and for a non-existent product costs none). Neither is severe; both are unnecessary leaks.

A second leak is presentational: the failure message is "Security code is not matching: carefully copy and do not change the link." — phrasing addressed to a human who mistyped a share link, not to the client that generated the link. The client can only emit a valid link, so this sentence can never be true for the client. It is either copied from a different product (one with shareable links) or written for the analyst who is, by now, reading the response body. On the evidence of §11.6 the former is likely: the sentence fits a click-to-download file host, not a machine-generated installer.

5.5 Identifier semantics

Three distinct identifiers circulate. They are often conflated; they are not interchangeable.

| Name | Example | Length | Where issued | Lifetime | Nature | |---|---|---|---|---:|---|---| | product id | 85892839 | 8 dec | service.php row key | until product deleted | server-side catalogue key; appears in the filename, the descriptor, and the signed download URL | | installation id | 3482d239f6935cd | 16 hex | GET /geo/ | one bootstrap | server-issued opaque token; 64-bit; used as TAP uid= and echoed as JSON userID | | channel / userID nonce | fbLzx, ByQX | 5 alnum | each getDownloadInfo response | one response | attribution/session nonce; regenerated per request; not persisted |

The distinction matters for privacy assessment:

  • the installation id is not a fingerprint. It is a random-looking value issued by the server, so its entropy comes from the server, not from the host: reinstall Windows and it changes; clear it from memory and it changes. Nothing binds it to hardware, account, locale, time zone or IP, other than the fact that the server can, if it logs the request, correlate the id with the requesting IP and clock. In GDPR terms it is a pseudonymous online identifier, not a device identifier.
  • the channel and userID nonces change on every request, so they cannot be used for cross-request correlation by the client; they can be used for correlation by the server only through its own log of (request time, IP, id, nonce). Their presence is therefore evidence that the server does keep such logs — otherwise the token would be pointless.
  • no other identifier is transmitted. No locale, no time zone, no screen metrics, no OS build, no install path, no user name, no machine name, no volume serial, no MAC, no SID, no tenant, no architecture, no language. The complete list of outbound scalar values is in §9.1.

This is, deliberately or not, a data-minimising design — and the single most defensible thing that can be said about the wrapper's handling of personal data.

5.6 Hosting and topology

Host A/AAAA Reverse / operator TLS Notes
dlsft.com Cloudflare 104.21.x, 172.67.x, 2606:4700:303x:: Cloudflare yes API host; server: cloudflare, alt-svc: h3
www.dlsft.com Cloudflare Cloudflare yes bootstrap /geo/
filedm.com Cloudflare Cloudflare yes brand, legal pages, analytics
directfiledl.com 116.202.13.47, 2a01:4f8:c012:42ca::1 static.47.13.202.116.clients.your-server.de — Hetzner no single host, plain HTTP, Express
dlsft.com/callback.php etc. Cloudflare — yes —
securestudies.com — — no (TAP) / yes (DPD) third-party measurement

The asymmetry is the finding:

  • the control plane (offers, descriptors, legal, telemetry) is on Cloudflare: TLS, HTTP/2, HTTP/3, anycast, DDoS absorption, origin hiding, and — significantly — the ability to serve HTTPS on names the operator does not have certificates for;
  • the data plane (the 16.7 MB archive) is on a single Hetzner cloud instance in Germany, reached by plain HTTP on port 80, with a Hetzner PTR record that names the hosting provider and nothing about the customer, and with no listener on 443.

server: cloudflare, cf-cache-status: DYNAMIC, cf-ray: …-CPH (Copenhagen edge) on the control plane; server absent (Express's default posture) on the data plane. The two planes are owned by the same commercial entity (§8) and are operationally separate, which is a reasonable small-operations design and an unreasonable design if the archive is meant to be trusted, because the only thing that can be authenticated about a plain-HTTP fetch is the bytes the middle of the path chose to deliver.

5.7 Assets referenced

Asset Referenced by Retrieved Content
dlsft.com/PO.jpg BDV DEFAULT_CONFIG 200, 32,755 B, image/jpeg, ETag "4377e68b…" 413 × 30 progressive JPEG, Adobe Photoshop CC 2019 (Macintosh), created 2019‑10‑04 08:01:55 −04:00, saved 08:30:04, sRGB IEC61966-2.1, photoshop:ColorMode=3
filedm.com/privacy.php BDV DEFAULT_CONFIG 200, 9 235 B "Privacy Policy", brand Filedm, quick-website.css, @fortawesome/fontawesome-free, .panelWidth{max-width:800px}
filedm.com/terms.php BDV DEFAULT_CONFIG 200, 7 635 B "Terms of Service", same template, <!-- Preloader -->
filedm.com/assets/img/brand/logo.png product descriptor 200 brand mark used in the wizard header
www.honeygain.com Glory (via SDK) 200 —
raw.githubusercontent.com/ShadowWhisperer/BlockLists/master/Lists/Ads AdReject 200 managed domain blocklist, plain text

The PO.jpg metadata is the most analytically productive single asset in the chain (it is 32 KB of JPEG and one sentence of provenance). Its embedded XMP history reads:

action="created"   when="2019-10-04T08:01:55-04:00"  softwareAgent="Adobe Photoshop CC 2019 (Macintosh)"
action="converted" parameters="from image/gif to image/jpeg"
action="saved"     when="2019-10-04T08:30:04-04:00"  softwareAgent="Adobe Photoshop CC 2019 (Macintosh)"

So the creative is a flattened GIF from October 2019, produced on a Macintosh, 34 days in under the thirteen-minute window of its own creation, and never touched again. It is also 413 × 30 — a strip, not a hero — i.e. the shape of an in-line banner slot in a ten-year-old offer wall, not of the 600 × 200 hero that a 2026 download manager would use. This single asset dates the interface template it belongs to earlier than the nlohmann::json version that serialises it, earlier than the Monaco build that sits beside it, and earlier than every one of the products it is used to sell.

5.8 Assessment

A‑5.1 — The control plane is profiled, the data plane is not. Cloudflare termination on every machine-readable endpoint, Hetzner static.*.clients.your-server.de on the one endpoint that serves the large object. This is a cost decision (bandwidth egress is cheapest at a bare host) that has a security consequence (the large object is the only one without TLS).

A‑5.2 — The s scheme is functional but misnamed. It is presented as a "security code"; it is an anti-crawl mask with no confidentiality, no integrity in the cryptographic sense and no freshness. Its one security-relevant property is accidental: because the algorithm lives in the client, it proves that the client is the Subject's build family — a weak client attestation. An unknown requester that supplies a valid s has either reversed the same 1 KB of code or copied a link, and the server cannot tell which. (Confidence: High.)

A‑5.3 — Identifier design is privacy-positive, incidentally. A server-issued 64-bit installation id with no host-derived inputs is the minimum viable attribution token. Whether this reflects intent or merely the absence of a hardware-id feature is unknowable from the binary, but the effect is measurable and is recorded in §9.2. (Confidence: High for effect; Low for intent.)

A‑5.4 — No defence in depth on the data plane. A plain-HTTP fetch of the principal artefact, with no hash, no length and no signature check on the client side, means that everything the product "is" at run time is asserted by one Hetzner instance and believed by one process. This is developed as F‑09 and rated in §10. (Confidence: High.)


6. The delivered artefact

6.1 Product descriptor

POST https://dlsft.com/service.php?request=getDownloadInfo&name=ByteBreaker V1_85892839 returns, verbatim (232 bytes, text/html, JSON body):

{
  "id": "85892839",
  "channel": "fbLzx",
  "userID": "ByQX",
  "title": "ByteBreaker V1",
  "description": "The ads shown in the installer help this exploit have no key system. For any issues ask on the discord server\t",
  "filename": "bytebreakerV1.zip",
  "logo": "https://filedm.com/assets/img/brand/logo.png",
  "payload": "http://directfiledl.com/getfile"
}

The description field is the only prose authored by the publisher in the entire record. It states the distribution model from the inside: the operator's installers show advertisements, the product therefore carries no key system, and support is provided on a Discord server. The trailing horizontal tab (0x09) is unescaped and survives JSON round-trip; it is a copy-paste artefact from a text field and is a small, useful fingerprint of a hand-entered row rather than a generated one.

The payload field is a base URL only — no query, no filename. The client appends ?id=<id>&s=<crc> (§5.4). This is why the URL is not discoverable by crawling filedm.com: the downloadable address exists only inside the client's arithmetic.

6.2 Catalogue properties

701 sequential identifiers (85892450–85893150) were probed with a constant prefix and a varying suffix. Exactly one resolved to a product: 85892839.

Supplementary probes:

Supplied name Returned title Returned id Returned filename
ByteBreaker V1_85892839 ByteBreaker V1 85892839 bytebreakerV1.zip
WhateverName_85892839 ByteBreaker V1 85892839 bytebreakerV1.zip
Foo_85892839 ByteBreaker V1 85892839 bytebreakerV1.zip
totally different_85892839 ByteBreaker V1 85892839 bytebreakerV1.zip
ByteBreaker V2_85892839 ByteBreaker V1 85892839 bytebreakerV1.zip
ByteBreaker V1 (no suffix) (default) Unzipper – 7zip.exe
85892839 (bare) (default) Unzipper – 7zip.exe
probe_99999999 — — "not valid id" (bare JSON string)
x_85892838, x_85892840 — – "not valid id"

Inferences, in descending confidence:

  1. The name is not the key; the suffix is. Renaming the prefix changes nothing; removing the suffix changes everything (falls to the default). (High.)
  2. The identifier space is global to the platform and sparsely populated. The subject's neighbours are unassigned, so the identifier is not a per-product sequence and not a content hash. It is an upload/row key allocated across all customers. (High.)
  3. There is a default product, returned whenever the name does not resolve: title Unzipper, description "7zip helps you unzip your files.", filename 7zip.exe, payload https://www.7-zip.org/a/7z2409.exe, logo the 7‑Zip logo. The default is benign, well-known, digitally signed, and — critically — retrievable without a key, which makes it a useful probe for whether a name resolves. (High.)
  4. ByteBreaker V2 does not exist in this identifier neighbourhood. The lookup resolved because the suffix was V1's; the prefix was ignored. A genuine V2 exists only if a different identifier does. (Moderate — it is not possible to prove absence over the whole space.)
  5. The default being "Unzipper" is a diagnostics artefact of some value: the no-match path returns a real, signed, third-party installer at a stable URL, which means an analyst can distinguish "server working, name wrong" from "server down" without any error message. (High.)

6.3 The archive

bytebreakerV1.zip
  size            16,787,263 B
  sha256          aa3fc2c089a8f039b2198d6f64e4e727e1631c7e18fe084429b1b491330ad85f
  compression     stored (method 0) for top-level entries; deflate within Editor/
  entries         289  (163 regular files, 126 directories)
  uncompressed    42,408,775 B
  ratio           2.53 : 1

stored on the executable and DLL entries is a positive finding: the executables are byte-identical to the files on disk before packaging, so archive extraction does not perturb image hashes and per-file hashing during triage is meaningful. It is also the signature of an assembly step performed by a tool that was told not to compress binaries (or by an operator adding to an existing archive).

6.4 File inventory

Path Size Date Role
bytebreakerV1.exe 5,667,926 2026‑05‑13 12:52 launcher / host (Themida, .NET 9)
QuorumAPI.dll 682,496 2026‑10‑27 16:50 managed bridge (184 types, 964 methods)
decompiler.exe 3,804,672 2026‑05‑12 16:23 Luau decompiler CLI (Medal, Rust)
version.json 30 2026‑10‑29 13:51 {"version":"1.4.816"}
lastKey.txt 16 2026‑05‑13 12:53 keysystemremoved
Newtonsoft.Json.dll 721,320 2025‑09‑16 10:04 JSON
Microsoft.Web.WebView2.{Core,WinForms,Wpf}.dll 603,712 / 38,976 / 82,000 2025‑01‑12 10:18 embedded browser host
runtimes/win-x64/native/WebView2Loader.dll 165,960 2025‑01‑12 10:18 WebView2 loader
bin/Quorum.dll 2,362,704 2026‑10‑20 22:46 engine (NativeAOT .NET 9, Luau)
bin/decompiler.exe 3,804,672 2026‑05‑12 16:23 duplicate
bin/libcurl.dll 652,800 2025‑12‑11 23:47 8.17.0‑DEV
bin/libssl-3-x64.dll 1,447,696 2025‑10‑01 19:02 OpenSSL 3.6.0
bin/libcrypto-3-x64.dll 7,428,376 2025‑10‑01 18:53 OpenSSL 3.6.0
bin/zlib1.dll 90,112 2025‑12‑11 23:21 1.3.1
bin/current_version.txt 5 2026‑10‑21 13:54 1.2.0
Editor/index.html 28,112 2026‑05‑13 12:35 WebView shell (brand)
Editor/Monaco/** ~15 MB 2024‑12‑27/29 vendored Monaco editor
Editor/info.evon 4 2024‑12‑27 true
Editor/theme.evon 129 2024‑12‑27 JSON theme manifest
Editor/tabs/New Tab 1 104 2024‑12‑27 scratch buffer
Editor/luaicon.png 17,025 2025‑01‑19 icon
Scripts/{Admin Script,Dex Explorer,Orca Hub}.lua 98 / 91 / 115 2026‑05‑13 loadstring stubs
autoexec/, workspace/, logs/ empty — runtime folders, never populated

The empty logs/, workspace/, autoexec/ directories are significant: the engine creates a timestamped log file in logs/ on every run (§6.7), and the editor writes a tab state into workspace/. Both are absent, and Editor/tabs/New Tab 1 still contains the publisher's scratch text. The archive is therefore a never-executed build, assembled by a tooling step rather than by a user session.

6.5 Identification

The highest-value single string in the deliverable is in the CodeView record of bin/Quorum.dll:

C:\Users\rizve\OneDrive\Desktop\Projects\Xeno\build\Release\net9.0-windows8.0\Xeno.pdb

This is a dotnet publish -c Release output path for a project named Xeno, target framework net9.0-windows8.0, built by a Windows profile named rizve. It is corroborated by:

Corroborator Detail
String census in bin/Quorum.dll Xeno ×59, xeno ×5, XENO ×3, Rizve ×9, rizve ×3
Contact literals in the engine rizve@xeno.onl, Discord handle .rizve
Product site literal https://xeno.now
CDN host x3no.pages.dev (leet form of xeno; pages.dev = Cloudflare Pages, i.e. a static site in the same account)
Runtime thread names Xeno Server thread: {:#x}, Xeno Caching thread: {:#x}
Environment contract %XENO_VERSION%, %XENO_SERVER_PORT%
Log strings [{} ({})][::BytecodePatchWatcherT], A Xeno resource was unable to load
Presence string "Using Xeno "
Upstream archive GitHub tyronetheqt/Xeno, M-r7z/Xeno-SRC — "Xeno: An external script executor for Roblox made entirely in C++", "credits Rizve", bundled 1.0.3 / 1.0.8 / 1.045

Finding F‑04 — The delivered executor is Xeno, and the "ByteBreaker" identity is a skin. The evidence is convergent from four independent directions (build path, in-engine self-identification, network naming, public source archive). Confidence: High.

6.6 Brand census and the cost of the rebrand

Artefact Xeno/xeno/XENO ByteBreaker/bytebreaker Quorum rizve Evon
bytebreakerV1.exe (Themida) 0 (literals encrypted) 0 (only in VS_VERSIONINFO) – – –
QuorumAPI.dll (managed) 0 0 23 – –
bin/Quorum.dll (engine) 59 + 5 + 3 0 0 12 –
Editor/index.html 0 4 + 1 – – –
Editor/theme.evon – – – – "ThemeManufacturer":"Evon"
version.json, lastKey.txt – – – – –

Read as a set, these columns describe a name-splitting exercise:

  • The engine says Xeno 67 times and ByteBreaker zero times. It is upstream, untouched, and it still advertises its own author (e-mail, Discord, website) to anyone who runs strings.
  • The launcher says ByteBreaker twice — both times inside the version resource, which Themida leaves in .rsrc because Windows Explorer reads it there.
  • The UI page says ByteBreaker five times, all in <title>, <h1>-class text, a Monaco theme name and the default script.
  • The bridge says Quorum twenty-three times and never says either of the other two.

So the product presents four names to four different audiences: Xeno to the reverse engineer, ByteBreaker to the user, Quorum to the developer reading the C# symbols, Evon to the theme system. Only ByteBreaker is new. It is applied in nine places: four HTML strings, two version resource fields, one window title, one assembly title, one file name.

The rebrand is complete, and it costs nine strings.

6.7 Component analysis — launcher (bytebreakerV1.exe)

Property Value Interpretation
Format PE32+ x64, GUI, 6 sections —
Section names empty, (8.0), (7.9), .imports(0.6), .rsrc(5.2), .boot(8.0) blank-name sections plus a .boot bootstrapper are characteristic of Themida / WinLicense (Oreans)
BSJB (CLR metadata root) absent metadata is encoded or relocated by the protector
mscoree / mscorlib / System.Private.CoreLib 0 / 0 / 0 CLR strings compressed away
Import table kernel32.dll × 1 every other import resolved at run time
TimeDateStamp 0xD1E3… → year 2081 deliberately invalid; a protector artefact
Code-section entropy 8.0 / 7.9 uniformly encrypted
Protector markers present Themida, Upx, uPK embeds a UPX-ish compression stage beneath Themida's VM
VS_VERSIONINFO ByteBreaker / ByteBreaker.exe / 1.0.0.0 / Copyright © 2024 survives packing (in .rsrc)

Finding F‑05 — The launcher is commercially protected. Confidence High. Consequence: string-based and import-based reasoning about bytebreakerV1.exe is unreliable; the managed API layer (QuorumAPI.dll) and the managed entry assembly are the tractable surfaces (§6.8, §6.9). The Copyright © 2024 and 1.0.0.0 fields indicate the launcher was built in 2024, roughly two years before the packaged engine — the product has been shipping with a frozen shell while the engine moved on.

The protector matters for two further reasons:

  • Detection surface. Themida rebuilds the import table, virtualises entry code, randomises section names per build, and (in trial/demo licensing) injects a licence module at run time. Static signatures built on .text bytes therefore have a shelf life measured in builds, and heuristic detections (HEUR/QVM, Gen:Variant, Packed.Win64.*) fire on the protector's shape rather than on the product's behaviour. A file can change its detection verdict while its logic is byte-identical.
  • Tooling surface. The .NET metadata the CLR needs at run time must be mapped at run time; during static analysis it may be absent, encrypted or relocated. dnfile/dnSpy therefore cannot open bytebreakerV1.exe without either a runtime dump or Themida-aware handling (§13.3), whereas they open QuorumAPI.dll without complaint.

6.8 Component analysis — managed bridge (QuorumAPI.dll)

Property Value
Size 682,496 B
Kind .NET assembly (BSJB at 0x092374)
TypeDef / MethodDef 184 / 964
AssemblyRef netstandard (2.0.0.0), Newtonsoft.Json
Sections 2 ("" + .rsrc) — consistent with a ReadyToRun/crossgen or single-file-merged image
Native imports none (normal for a managed image)
Obfuscation none observed: namespace, type, method, property and field names are all readable
Version resource CompanyName QuorumAPI, ProductName QuorumAPI, 1.0.0.0, empty copyright

Recovered identifier inventory (BCL names removed; these are this application's names):

AttachAPI   Attach   Execute   IsAttached   IsRobloxOpen   IsAutoInjectEnabled
SetAutoexecPath   AutoUpdate   IsApplicationTrustedToRun   Authenticode
scriptSource   get_PayloadNames   get_UrlsForUri   get_Theme   get_Enabled
get_SystemTag  get_WarningTag  get_InfoTag  get_ErrorTag  get_SuccessTag
get_Red  get_Green  get_White  get_Orange  get_LimeGreen
get_JoeBiden   get_JoeBiden4
MachineKeysetAlgorithm   SameMachine   FileShare   StrongNamePublicKeyBlob
UnauthenticatedPrincipal  Appendix…  A6A_N_K0AEAVDataSink  LicenseFile  IsWellKnownEqualityComparer

Four things may be read directly out of that list, without any decompilation:

  1. The public API surface of the executor is these names. Attach(), Execute(), IsRobloxOpen(), IsAttached(), SetAutoexecPath(), AutoUpdate(), IsAutoInjectEnabled(). This is a product specification accidentally shipped: the wrapper can attach to a running client, auto-inject on process start, set an auto-execute directory, and self-update.
  2. A trust/authenticity subsystem exists. IsApplicationTrustedToRun, Authenticode, StrongNamePublicKeyBlob, StrongNameSignatureGeneration…, MachineKeysetAlgorithm, SameMachine, AuthenticationInstant, IsAuthenticated, KeyContainerPermissionFlags, IsSecurityCritical, DebugFacilityMessage, LicenseFile. Combined with SameMachine and MachineKeysetAlgorithm, this is the shape of a machine-bound licence check — a hardware- or machine-key-based validation, not a hardware fingerprint (§9.2, §9.4). The distinction is material and is drawn there.
  3. An auto-update subsystem exists (AutoUpdate, update-available, update-deferred, update-rolled-back, §7.3), and it is manifest/tag-driven, not a plain "download latest .exe".
  4. A theme system exists, and its colour members are Red, Green, White, Orange, LimeGreen, with semantic tags System, Warning, Info, Error, Success. The two joke members JoeBiden and JoeBiden4 sit inside the same type. (JoeBiden4 implies at least three predecessors were deleted.)

Finding F‑06 — The managed layer is unobfuscated and fully decompilable. Confidence High. Consequence: for any future sibling or updated build, QuorumAPI.dll is the cheapest possible entry point — it can be read as C# in ILSpy in minutes, whereas the engine requires NativeAOT-aware tooling. An analyst with one hour should spend it here.

6.9 Component analysis — engine (bin/Quorum.dll)

Property Value
Size / timestamp 2,362,704 B / 2026‑10‑20 20:45:16 UTC
Kind .NET 9 NativeAOT native image (net9.0-windows8.0)
Exports Attach, Execute, GetClients, Initialize, SetSetting, Version
Static runtime MSVC (msvcp140, vcruntime140, vcruntime140_1)
Dynamic deps libcurl 8.17.0‑DEV, libssl-3-x64/libcrypto-3-x64 OpenSSL 3.6.0, zlib1 1.3.1
System deps ws2_32, crypt32, advapi32, user32, shell32, dwmapi, UCRT -api-ms-win-crt-
Script VM Luau, statically linked (.?AVCompileError@Luau@@, .?AVParseError@Luau@@)
HTTP server cpp-httplib (header-only; base64 alphabet, Content-Disposition, multipart/form-data, localhost)
Crypto AES-GCM, SHA-256, HMAC, RSA, X.509
Compression zstd (encode + decode)
Console allocates one (CONIN$, CONOUT$)
Threads Xeno Server, WebSockets Cleaner, Xeno Caching, Client Checker
Environment %XENO_VERSION%, %XENO_SERVER_PORT%
Logging logs\<%Y-%m-%d_%H-%M-%S>.log
Contact strings rizve@xeno.onl, .rizve

The six exports are the product contract:

Initialize(env?)          start; bind the local HTTP port; load VBX/cache
Attach(pid|handle)        inject into a Roblox client and start the Luau environment
Execute(source, client)   run a script chunk in that client
GetClients()              enumerate live RBXClient objects (multi-instance)
SetSetting(key, value)    autoexec path, injection mode, RPC, theme, …
Version()                 version tuple for the UI and the update check

Finding F‑07 — The engine is a modern Luau host, not a Lua 5.1 wrapper. Confidence High. The Luau symbols present (LuauSolverV2, LuauTypeLengthLimit, LuauStoreConstKeywordBegin, LuauVirtualBcBuilder, LuauIntegerType2, LuauNoDuplicateBinaryPrefix, LuauExportValueSyntax, LuauCompileIifeInline, LuauIntegerFastcalls, LuauBytecodeCostModel) are internal, non-public feature-gate identifiers from the Roblox/luau compiler, not documentation terms. They date the linked revision to a recent upstream and prove the VM is not LuaJIT, not Lua 5.1, and not a re-skin.

Consequences for the user: typeof, task.*, buffer, string interpolation, generalized iteration, vector/normal types and Luau type syntax all behave natively, because they are native. This is the single strongest quality signal in the entire deliverable, and it is independent of the vendor's marketing.

6.10 Component analysis — decompiler (decompiler.exe)

size            3,804,672 B      date 2026-05-12
language        Rust              rustc x75, panic/backtrace machinery, std static
crate/CLI       "medal", "medalCli"
behaviour       Decompile <input> <output> [encode] ; luaudecompile
bytecode target Luau / Lua 5.1 family
argument parser clap (github.com/clap-rs/clap)
imports         kernel32, ws2_32, ntdll, bcryptprimitives, vcruntime140, api-ms-win-crt-*

Finding F‑10 — The shipped decompiler is Medal, an open-source Rust Luau decompiler, invoked as a child CLI. Confidence High. Two operational consequences:

  • A .lua file that the user "decompiles" in the editor is reconstructed by a separate process, driven by CommandLine arguments, which means the temporary input/output paths appear on disk and in process-creation telemetry. They are removed by the host (the deleted-temp-file pattern appears in logs/ and in process-exit behaviour), but they exist for the duration.
  • The ws2_32 import set (18 functions) in a decompiler is the Rust standard library's socket scaffolding — not a feature. It is a useful reminder that "why does this CLI import the network stack" is normally answered by "because Rust does", and that imports are a poor proxy for behaviour in modern statically linked binaries.

6.11 Component analysis — editor (Editor/) and bundled scripts

The editor is a vendored Monaco (VS Code) build dated 2024‑12‑27/29, served from Editor/index.html (28,112 B), with a custom theme ByteBreakerTheme (base: vs-dark), a localStorage schema key bytebreaker.editor.tabs.v2, and a default document:

-- Welcome to ByteBreaker :D
-- Have Fun!

print('Hello World')

Editor/theme.evon (129 B) is a JSON theme manifest whose only manufacturer entry is:

{ "Theme": [ { "ThemeManufacturer": "Evon", "Color1": "#FF9700FF", "TextboxImage": "" } ] }

and Editor/info.evon is the four-byte string true. The .evon extension is otherwise unattested; it is the only self-authored artefact format in the deliverable. It is also the only artefact that names a party other than Xeno/rizve/AN Soft — and §8 identifies "Evon" as the party that performed the re-brand.

Editor/tabs/New Tab 1 (104 B) preserves the publisher's scratch buffer:

loadstring(game:HttpGet("https://raw.githubusercontent.com/Lucasfin000/SpaceHub/main/SpaceHubLoader"))()

a third-party hub loader, unrelated to Xeno, left behind by the person who checked that the editor worked.

Scripts/ contains three loadstring stubs, each fetched from a moving branch at run time:

File Upstream Note
Admin Script.lua EdgeIY/infiniteyield → master/source canonical Infinite Yield; branch, not tag
Dex Explorer.lua infyiff/backup → main/dex.lua a mirror account, not the canonical Dex repo
Orca Hub.lua richie0866/orca → master/public/latest.lua CRLF line endings ⇒ authored on Windows

"The bundled script library" is therefore three live GitHub URLs, none pinned to a commit or a tag. Whoever controls a branch controls code that runs in the executor's Luau environment with executor privilege. This is normal for the genre and is a supply-chain consideration, not a defect: pinning would staleness the scripts, not pinning means the last writer to the branch wins.

6.12 Component analysis — library versions

Library Version Release date Observation
curl 8.17.0‑DEV 2025‑12‑11 build a development snapshot, not a tagged 8.x release
OpenSSL 3.6.0 2025‑10‑01 build matches the 3.6 series released 2025‑10
zlib 1.3.1 2025‑12‑11 build current stable as of 2025
WebView2 SDK 1.0.2xxx (2025‑01‑12 file date) 2025‑01‑12 ~1.0.3110 era
Newtonsoft.Json 13.x (721,320 B, 2025‑09‑16) 2025 current 13.0.3-era size
Monaco 0.5x (2024‑12‑27) Dec 2024 VS Code 1.9x contributor build
Sciter 19,290,112 B sciter.dll (Glory, not the executor) — Terra Informatica HTML UI engine
nlohmann/json 3.11.2 2022‑08 the only older library in the set
RapidXML 1.x (header-only) — no version string

Finding F‑11 — The author tracks upstream. Confidence High. OpenSSL 3.6.0, zlib 1.3.1 and a curl 8.17.0 development snapshot are all contemporaneous with the build, not with the project's inception. That is the opposite of the "vendor drops in a 2011 OpenSSL and forgets it" pattern that dominates cheap Windows software. The single stale dependency is nlohmann/json 3.11.2, which is a header-only library that does not need attention, and the oldest licensed component is Newtonsoft.Json (required by the WebView2 managed API).

Library hygiene is not a safety verdict (a careful dependency set can still fetch an unsafe script), but it is evidence of ongoing maintenance rather than abandonment, and it makes the "abandoned rebrand that will never be patched" hypothesis (hypothesis H‑2, §10.5) less likely, at least for the executor.

6.13 The remote configuration object

At first attach the engine fetches a small binary from the network and keeps it in memory for the lifetime of the process. Two sources are configured; they are tried in order and the second is used only if the first fails. Both were retrieved and compared.

Property Primary Secondary
URL https://pastebin.com/raw/A7n0vJtR https://x3no.pages.dev/ep.zst.enc
Hosting pastebin.com (US), Cloudflare-fronted Cloudflare Pages (static)
Content-Type text/plain; charset=utf-8 application/octet-stream
Transfer base64 text, 344 characters binary
Wire size 344 B 276 B
Decoded size 276 B 276 B
Cache-Control public, max-age=14400 (4 h) default
SHA-256 e4b57255b6b01b12a7a729d3155294600ef2d43af2ae13d6e1126f15a4126381 (decoded) e4b57255b6b01b12a7a729d3155294600ef2d43af2ae13d6e1126f15a4126381
Byte-for-byte identity 276 / 276 equal, 0 differing —

Finding F‑08 — The two configured sources are mirrors of one object, not a fallback pair. Confidence High. The values are byte-identical after decoding. The "Primary/Secondary" control flow and its error strings are therefore host-redundancy: one copy optimised for availability behind a paste service that most networks allow, one copy served as a raw static object from the project's own CDN. Nothing rotates, nothing differs, and nothing cross-checks — which is the one design property worth criticising here: a corrupted or stale mirror is accepted by the client (§6.7, F‑09).

6.13.1 Envelope decomposition

276 bytes are divisible in exactly one natural way for AEAD:

276 = 12 + 248 + 16
      |    |     └── 128-bit authentication tag
      │    └──────── 248-byte ciphertext
      └─────────────  96-bit nonce   (the ubiquitous GCM nonce size)

and 248 bytes is not divisible by blocksize for any block cipher in ECB mode (248 mod 16 = 8), so the payload is not a padded block-mode encryption — it is a stream construction (CTR/GCM/ChaCha-like). Combining the two:

                    ┌──────────────────────────── 276 B ────────────────────────────┐
                    │  12 B nonce   │        248 B ciphertext        │   16 B tag        │
                    │               │       (⊕ keystream)            │                   │
    plaintext   =   │               │   zstd frame  →  JSON           │                   │
                    └──────────────────────────────────────────────────────────────────┘

  decompressed size estimate (ciphertext 248 B, zstd ratio ≈ 0.45 .. 0.60 for short JSON) ≈ 400–550 B

Supporting observations:

  • the whole blob is high-entropy (7.195 bits/byte over 276 bytes) with no zstd magic 28 B5 2F FD at any offset, no gzip 1F 8B, no PK, no XZ — so compression is inside the cipher, i.e. the pipeline is encrypt(compress(x)), which is also the correct order;
  • the engine carries AES, GCM, nonce, Encrypt/Decrypt and zstd/decompress as literals;
  • the pinned sizes (12 nonce, 16 tag) are fixed literals rather than parameters — fixed as a compile-time constant generates a single hex string (e.g. "12", "11…"), not a parse; the presence of the constants in this form is the reason for the moderate confidence on layout rather than high;
  • the alternative 16-nonce / 12-tag pairing was tested and also fails (§6.14) — which is consistent with the pinned layout, not proof of it.

Confidence on layout 12‖248‖16: High (length arithmetic + entropy + crypto import set). Confidence on algorithm AES-GCM: Moderate‑High (AES/GCM literals present, but ChaCha20‑Poly1305 is also 12+16 and the literal ChaCha/Poly1305 were not found — see §6.14, which is why GCM is the preferred reading).

6.14 The key sweep (and what it rules out)

If the object is AEAD under a fixed key, the key must exist somewhere in the engine. bin/Quorum.dll is 2,362,704 bytes. The hypothesis "the key is a contiguous byte window of the image" was tested exhaustively:

for klen in (32, 16):                       # AES-256 and AES-128
    for off in range(0, len(image) - klen):  # every offset, step 1, both alignments
        k = image[off : off+klen]
        if k == all-zeroes: continue
        for blob in (primary, secondary):
            for layout in (nonce-first 12‖C‖T, nonce-first 12‖T‖C):
                try AESGCM(k).decrypt(nonce, ct+tag, None)
                if success: report(off, klen, blob, layout, plaintext)

Result:

key candidates evaluated   9,366,416
algorithm                  AES-128-GCM, AES-256-GCM
layouts                    2 (tag appended; tag prepended)
blobs                      2 (byte-identical mirrors)
elapsed                    26 s
successful decryptions     0

Finding F‑09 — The configuration key is derived, not stored. Confidence High for the negative; the corollary (that it is derived at runtime) is by elimination, Moderate‑High. This is a small but genuinely useful result, because it constrains where to look for the plaintext:

  1. The key is not a literal array → it is a function of something: a passphrase, a build constant, the product id, the installation id, a machine value, secretA/secretB (§6.15), or a constant in the launcher (bytebreakerV1.exe, not swept — Themida).
  2. The nonce is not in the plaintext-recoverable set either: if the key were derived from the nonce, the sweep would have found a fixed-point pair. It found none, so nonce and key are independent inputs.
  3. A 16- or 32-byte constant with an unusual value anywhere in the image was excluded as the key. The remaining candidates are, in order of likelihood: an unaligned window in a compressed resource (EVP_…/sciter tables), a PBKDF/HKDF output seeded by a detectable string, or an XOR-folded constant computed in a dead-cold path that a static sweep cannot linearise past.

The practical significance is not cryptographic. AES-GCM with a derived key and a random nonce is adequate; the failure mode is not cryptanalytic, it is operational — the key must be derived from something that exists on every machine, so it exists on every machine.

6.15 The third secret pair

The engine derives, or holds references to, two additional short strings whose literals appear in the frozen string segment adjacent to the presence block (§6.7) but which have no recoverable consumer in the analysed build:

RVA 0x…A0D48   "xeno"
RVA 0x…A0D60   "secretB"
RVA 0x…A0D68   "secretA"
RVA 0x…A0D70   "https://xeno.now"
RVA 0x…A0DA8   "icon"
RVA 0x…A0DC0   "Using Xeno "
RVA 0x…A0DD8   "1393297926209405000"

secretA and secretB are field names, not values. In the Discord RPC activity schema, secrets is an object carrying exactly two semantically relevant members for executable games:

"secrets": { "join": "<opaque>", "spectate": "<opaque>", "match": "<opaque>" }

and the engine's own event strings cover the join lifecycle:

"Connected to {}, userId: {}"
"Disconnected ({}): {}"
"{}: {}"
"You joined: {}"
"{} requested to join"
"RPC is Enabled"
"RPC was disabled"

Finding F‑13 — The engine publishes a Discord Rich Presence with join/spectate secrets and an inbound join-request handler, bound to the Roblox userId and placeId. Confidence Moderate‑High. The secrets object and the event strings are direct; the mapping of secretA/secretB onto join/spectate is inferred from field order and from the join-lifecycle strings, and is stated as an inference.

Analytic consequence: the engine has a two-way channel to the Discord client, not a display-only one. A join secret turns the presence card into a transport — a friend clicks Join, Discord's local service emits ACTIVITY_JOIN, the engine receives it and initiates a Roblox join. This is a feature, not a covert channel, but it is a feature in the wrong direction for a threat model that assumes "presence is display-only", and it explains a string inventory that a display-only presence would not need ("Requested to join", "You joined", "Disconnected ({})").

Two strings also survive from the feature's absence:

"RPC was disabled"           (explanatory text, shown when user has RPC off / opted out)
"WebSockets Cleaner thread"  (dormant; the WS client in this build does not open)

A cleaner thread whose socket never connects is the code smell of a switched-off feature (telemetry‑ or rollout‑limited), and is the strongest single indicator in this binary that the vendor plans to change the observable protocol without a file update — i.e. that the user's network fingerprint is not yet at its steady state.

6.16 The update contract

version.json (1.4.816), bin/current_version.txt (1.2.0) and the engine's %XENO_VERSION% environment variable form a three-way version model:

launcher (Themida)   reads  version.json         → current release of the product
packaged engine      reports 1.2.0                → the build shipped inside the archive
engine at runtime    exports %XENO_VERSION%       → the build that is actually executing

Finding F‑20 — The packaged build is a stale pairing and the client self-updates engine components at first run. Confidence Moderate‑High. The evidence (engine dated 2026‑10‑20, current_version.txt 2026‑10‑21, version.json 2026‑10‑29, archive metadata 2026‑10‑29) describes a packaging step that refreshed its manifest eight days after its binaries and, by the launcher's own AutoUpdate state machine (§6.8), reconciles at start-up. The direct test in §13.1 (two runs, ten days apart, compare logs/ and check for a network round-trip to a non-Xeno host) should be treated as pending.

The consequence is stated carefully, because it is easy to overstate:

  • Not "the archive is out of date" in the sense of decayed; the author is actively shipping (§6.12).
  • Yes: the bytesexecuted at run time are a strict superset of the bytes assessed here, and that superset is defined by a remote manifest that is mutable without authentication (§6.17).
  • The modulo is a small, stable, well-typed surface: the engine acquires a new bin/ from a manifest, verifies hashes (loopfor 256-bit 0x…), and executes it. An attacker cannot inject arbitrary code into the initial state, but they can inject arbitrary code into the subsequent state if they control the update origin, and there is no signature pinning in the analysed build (§6.17).

6.17 Trust boundaries in the update model

  LAUNCHER (signed log: AN Soft, 2026-11-17)  ─┐
                                              ├─ statically trusts
  version.json  (inside the archive)          ─┤   nothing but the archive itself
                                              │
  manifest (latest.manifest) ──► update zip ──┘ ──── trusted only by
                 │                          ───────────  hash match
                 │                          (sha-256 per file, in the manifest)
                 ▼                                    │
            engine binaries                             ▼
            (published as plain HTTP? no: TLS on the     signature check: NONE
             update host — see §5.6)                     hash check:     YES
                                                         rollback:       YES
                                                         expiry:         NO

The dots that are actually joined, and their consequences, are set out in §10.3 as three fault lines.


7. The offer bundle

7.1 Composition

The feed returned five offers to a United Kingdom egress on 2026‑11‑16. Across nine channel values tested the set was invariant; only the reflected echo length varied, confirming that channel is attribution metadata and not a targeting key (§7.6).

# id Display name Category Publisher Delivery
1 operagx OperaGX mainstream browser Opera Norway AS (affiliate) CPA redirect
2 glory Glory gameplay capture + bandwidth AN Soft (first party) Velopack manifest → ZIP
3 adreject AdReject ad blocker + bandwidth AN Soft (first party) Velopack manifest → ZIP
4 aichat AIChat AI chat client + bandwidth third party (Tauri/Rust) GitHub release (NSIS)
5 po Premier Opinion market-research panel comScore / VoiceFive direct EXE

Three of five are bandwidth-sharing clients. Two of those three are first-party products of the company that built the wrapper. One is a genuine mainstream product. One is a survey panel.

7.2 Offer 1 — operagx (Opera GX)

{ "id": "operagx", "name": "OperaGX",
  "link": "https://www.ovardu.com/6R9H6BD/KMZXBTT/?sub1=FileDM&sub=default",
  "params": "--silent --allusers=0",
  "targetGeos": [], "fileChecks": [],
  "registryChecks": [
    {"root":"HKEY_CURRENT_USER","path":"SOFTWARE\\Classes\\OperaGXStable","value":"URL Protocol"},
    {"root":"HKEY_LOCAL_MACHINE","path":"SOFTWARE\\Classes\\OperaGXStable","value":"URL Protocol"},
    {"root":"HKEY_CURRENT_USER","path":"SOFTWARE\\Classes\\OperaStable","value":"URL Protocol"},
    {"root":"HKEY_LOCAL_MACHINE","path":"SOFTWARE\\Classes\\OperaStable","value":"URL Protocol"},
    {"root":"HKEY_CURRENT_USER","path":"SOFTWARE\\OperaGX","value":"installed"} ],
  "privacy": "https://legal.opera.com/privacy/",
  "eula":    "https://legal.opera.com/eula/computers/" }

Observations:

  • The link is not an Opera URL. It is ovardu.com/6R9H6BD/KMZXBTT/, a redirect slug, carrying sub1=FileDM&sub=default. sub1 names the publisher; sub is the placement. This is standard affiliate-tracker grammar (HasOffers/Voluum/Cake lineage) and means the payout is counted by the tracker, not by Opera — i.e. an intermediary network sits between AN Soft and Opera, and takes a margin.
  • Five suppression keys, two brands, two hives. The offer is hidden if the user already has Opera GX or Opera, installed per-user or per-machine, or if a native SOFTWARE\OperaGX\installed key exists. That is a knowledgeable affiliate: duplicate installs are the main chargeback cause and this configuration is engineered to prevent them.
  • --allusers=0 is the only scope switch in the bundle (§4.5).
  • targetGeos empty ⇒ offered worldwide, which is correct for a browser with a global payout schedule.
  • The privacy/eula fields point at Opera's real legal pages; the offer page therefore links a vendor whose document set exists, which is a small but real reputational transfer.

7.3 Offer 2 — glory (Glory)

{ "id": "glory", "name": "Glory",
  "link": "https://glory.tv/download",
  "params": "--silent --launch --channel IU",
  "targetGeos": ["!DE"], "registryChecks": [], "fileChecks": ["%ProgramFiles%\\Glory"],
  "privacy": "https://glory.tv/privacy", "eula": "https://glory.tv/terms" }

Fulfilment chain:

GET https://glory.tv/download                                  (200, web installer, AN Soft-signed)
  └─ GET https://glory.tv/releases/windows-x64/latest.manifest  (200, 1 077 B, plain-text key=value)
       product=Glory  channel=stable  version=1.0.7  architecture=x64
       package_url=https://glory.tv/releases/windows-x64/Glory-1.0.7-windows-x64.zip
       package_size=12579642
       package_sha256=1b67e6b6…93263101c
       file_count=3
       file=Glory.exe|2167688|3111a67c…aab62c162525068
       file=hgsdk.dll|7765672|2720d215…186da186850a37c6c
       file=sciter.dll|19290112|8ec04a9f…f4ab87ed8b883270

The downloaded archive verified exactly against all four hashes advertised in the manifest.

Contents, after extraction:

File Size Identity
Glory.exe 2,167,688 AN Soft; FileDescription Glory; 1.0.7; PDB C:\Users\Samim\Desktop\Glory\x64\Release\Glory.pdb
hgsdk.dll 7,765,672 Honeygain, UAB; Honeygain SDK service; 1.5.0; DigiCert OV
sciter.dll 19,290,112 Terra Informatica Sciter HTML/CSS/script UI runtime

Glory.exe imports (mfplat, winhttp, crypt32, ole32, shell32, d3d/D2D class) and its WinRT-generated headers identify it as a Windows.Graphics.Capture application: it records a window or monitor through the DWM, not the desktop through GDI. PDB names preserved in the image: Recorder.cpp, ScreenshotCapture.cpp, Generated Files\winrt\{base,Windows.Foundation,Windows.Graphics.Capture}.h.

The SQLite schema (§8.2) confirms that recordings and screenshots are catalogued and uploaded, with server_media_id and share_url returned by a backend.

Assessment. Glory is a shareware gameplay-capture application — the Medal/Outplayed/ShadowPlay class — monetised not by subscription but by proxyware. Its iconography and copy describe it as "Never miss a great gaming moment", which is accurate: it does record. The bandwidth-sharing sentence is present in the offer description (§7.9) and, per the vendor's own settings defaults, enabled.

7.4 Offer 3 — adreject (AdReject)

{ "id": "adreject", "name": "AdReject",
  "link": "https://adreject.com/download",
  "params": "--silent --channel IU",
  "targetGeos": ["!DE"], "registryChecks": [], "fileChecks": ["%ProgramFiles%\\AdReject"] }
GET https://adreject.com/releases/windows-x64/latest.manifest
  product=AdReject  version=1.0.9  architecture=x64
  package_url=https://adreject.com/releases/windows-x64/AdReject-1.0.9-windows-x64.zip
  package_size=574773  package_sha256=59358bb5…acba726e0a  file_count=5

AdReject-1.0.9-windows-x64.zip  (574,773 B, sha256 59358bb5…, verified)
  AdReject.exe          907,144   e943c047…  AN Soft 1.0.9  © 2026
  WinDivert.dll          47,616   c1e060ee…  (unsigned in package)
  WinDivert64.sys        94,144   8da08533…  (unsigned in package)
  licenses/WinDivert-LICENSE.txt  61,349
  licenses/miniz-LICENSE.txt       1,184

AdReject.exe imports: d2d1, dwrite, dwmapi (Direct2D/DirectWrite UI — no Sciter here), advapi32 (25), bcrypt (11), crypt32, shell32, ole32, oleaut32, user32 (54).

The binary contains its own command-line documentation, which is unusual and useful:

AdReject.exe --console [blocklist]
AdReject.exe --update-lists
AdReject.exe --enable | --disable
AdReject.exe --pause <seconds> | --resume
AdReject.exe --update

and the accompanying strings:

Invalid pause duration; use 0 through 86400 seconds.
--pause requires a duration in seconds (maximum 86400).
Managed-list update failed:
Managed ad list updated:
Managed list size changed by more than the safety limit (
Managed-list download exceeded the 32 MiB safety limit
managed list update failed; retaining last known-good list:
blocklists reloaded after a file change
blocklist reload rejected; retaining current rules:
Blocklists contain no usable domains
is BLOCKED by lists/blocklist.txt.
is not present in an active blocklist.
Global\AdReject.Update.Request.v1
Unable to create control-channel security descriptor: Windows error
"bandwidthEnabled":true            "bandwidthEnabled":false
bandwidth-on  bandwidth-off  consent
DiagnosticsEnabled  PendingUpdateTelemetry
https://raw.githubusercontent.com/ShadowWhisperer/BlockLists/master/Lists/Ads

Points of substance:

  • Kernel driver, user-mode policy. The ad blocker is not an extension and not a hosts file; it is a user-mode JSON policy driving WinDivert, an LGPL Windows packet-divert/WFP driver that hands userland both inspection and re-injection. From first launch until termination, the driver is loaded on the machine and the user-mode process holds the handle. The consequence (§10.2) is that the product's networking risk is not proportional to its ad-blocking value: a 907 KB binary with a 90 KB driver sits in the packet path of every process on the host.
  • Bounded operation. --pause is clamped to 0 … 86400 (24 h) with a typed error, and --resume exists. A pause that was offered but unbounded, or clamped to a century, is the common defect; this is the correct one, and it means a deliberate, time-scoped absence of filtering is representable.
  • Two-channel control with a security descriptor. --update-lists from a shortcut and the UI's own update path must agree, so the process exposes a named object (Global\AdReject.Update.Request.v1) and — note — explicitly sets a security descriptor on it ("Unable to create control-channel security descriptor: Windows error"). A Global\ object created without a DACL is writable by any user in the session on some configurations; the code author knew this.
  • Update integrity rails. A 32 MiB size ceiling on the managed list; a relative "size changed by more than the safety limit" guard against a list that doubles overnight; and retaining last known-good list / update-rolled-back semantics. A failed or suspicious update does not leave the machine with no list — it leaves the previous list.
  • Telemetry-free by design, with an opt-in. DiagnosticsEnabled and PendingUpdateTelemetry are settings, not channels; the update outcome telemetry (update-available, update-downloaded, update-install-started, update-succeeded, update-failed, update-rolled-back, updater-compatibility, updater-launch) is evented locally and uploaded only if diagnostics are on. This is a better-designed update path than the wrapper that installs it.
  • The blocklist source is a third-party GitHub account. ShadowWhisperer/BlockLists publishes categorised domain lists (ads, tracking, malware, adult, gambling, fakenews, social) as plain text for Pi-hole/AdGuard/NextDNS-style consumers. AdReject therefore ships zero rules of its own and depends on one individual's master branch — a single point of influence over what the product blocks, and (because master, not a tag) no protection against an unexpected change of content beyond the size guard.

7.5 Offer 4 — aichat (AIChat)

{ "id": "aichat", "name": "AIChat",
  "link": "https://github.com/aichatdl/aichat-app-tauri/releases/download/v2.0.3/aichat_2.0.3_x64-setup.exe",
  "params": "/S",
  "targetGeos": ["DE","AT","CH","IT","FR","ES","GB"],
  "registryChecks": [ {"root":"HKEY_CURRENT_USER","path":"SOFTWARE\\AIChat","value":"installed"} ] }
aichat_2.0.3_x64-setup.exe   3,807,352 B   sha256 6430a772eb047467aa0c3e4bc964719dbf85b9adc58e63738a8652099ba6f8ec
  format    Nullsoft Scriptable Install System (NSIS) self-extracting
  payload   $PLUGINSDIR/nsis_tauri_utils.dll, NSISdl.dll, System.dll, modern-header.bmp, aichat.exe (13,702,656 B)
  runtime   Tauri 2.x (Rust) + WebView2; minisign updater
  updater   https://github.com/aichatdl/aichat-app-tauri/releases/latest/download/latest.json
  pub_date  2026-06-11T13:10:36.323Z      version 2.0.3
  backend   http://my.aichat.rocks:9005   (note: cleartext, non-standard port)
  IPC       http://tauri.localhost · http://ipc.localhost · http://asset.localhost

Assessments:

  • The only offer that is fully open and hash-pinned by a third party. The installer is in a public GitHub release, the updater is minisign-signed and Tauri exposes the public key in the binary (RWShzQOk…, key B3ACBEE7A403CDA1), and the payload is reproducible from a version-controlled source tree. From a supply-chain standpoint it is the most verifiable item in the bundle.
  • The only offer that fetches its application bundle from GitHub and the only one whose update signing is independent of AN Soft. This is also why it is the only offer that could survive the disappearance of its publisher.
  • The only offer whose geography is a positive list (DE AT CH IT FR ES GB): a D‑A‑CH + EU‑5 launch footprint, consistent with a young product choosing markets it can support in one language set. Its overlap with the !DE rule of the two AN Soft products (§7.10) is the single most interesting fleet-level fact in the bundle.
  • NSISdl.dll + System.dll + nsis_tauri_utils.dll in $PLUGINSDIR are the stock plugin set of the tauri-action NSIS template. Their presence, and nothing else, is clean: NSISdl is present but unused (Tauri fetches at run time); it is a template artefact, not a capability.
  • http://my.aichat.rocks:9005 — cleartext, port 9005, on a domain whose :443 is Cloudflare-fronted. A Tauri application calling a plain-HTTP API on a fixed high port is calling a host-local service (a frontend the project expects to run on localhost:9005) or a VPS behind an A record with no certificate. Both readings are supported; the literal occurs once and no TLS variant is present.

7.6 Offer 5 — po (Premier Opinion)

{ "id": "po", "name": "Premier Opinion",
  "link": "https://post.securestudies.com/packages/PI1032/ContentI3.exe",
  "params": "-c:1538 -t:InstallUnion",
  "targetGeos": ["US","CA","GB","ES","IT","BR","IN","FR","AR","MX","HK","JP","DE","TW","NO","SG",
                 "IE","MY","AU","CH","NZ","CL","SE","FI","ID","VN","CO","PE"],
  "privacy": "-", "eula": "-",
  "fileChecks": ["%ProgramFiles%\\PremierOpinion"] }

Fulfilment: GET https://post.securestudies.com/packages/PI1032/ContentI3.exe — note the PI1032 path segment (a package identifier, presumably the panel installer's build number) and the -c:1538 -t:InstallUnion arguments which mirror the wrapper's own campaign and network identifiers (§7.8). The privacy and eula fields are literally "-" because the hyperlinks are embedded in the description field as HTML anchors:

<a href="https://www.premieropinion.com/common/termsofservice-v1">Terms of Service</a>
<a href="https://www.premieropinion.com/privacy-policy">Privacy Policy</a>

which is the documented reason the description is long (§7.9), and the reason the interface must render SysLink — a plain STATIC would show the markup.

ContentI3.exe was not retrieved in this engagement (OQ‑7). Its identity is corroborated by public sandbox reporting, which classifies it as a VoiceFive Networks-signed PremierOpinion installer that lays down pmropn.exe and related components.

7.7 Offer-by-offer comparison

Dimension OperaGX Glory AdReject AIChat PremierOpinion
Publisher Opera (NO) via tracker AN Soft (NL) AN Soft (NL) unknown (Tauri) comScore/VoiceFive (US)
Signed? yes yes, AN Soft yes, AN Soft unverified at byte level; NSIS + minisign updater yes (historically)
Installer class proprietary short WebSetup short WebSetup NSIS proprietary WebSetup
Managed payload yes no (C++/WinRT/Sciter) no (C++/D2D) Rust/Tauri yes (.NET)
Persistence browser default user user user panel + proxy config
Admin needed? no (--allusers=0) no yes (driver) no (/S) yes (HKLM)
Bandwidth sharing no yes (Honeygain SDK) yes (own) yes (own) no (measurement instead)
Records the screen no yes (DWM) no no no (measures browsing)
Kernel component no no yes (WinDivert64.sys) no no
Geographic rule worldwide worldwide − DE worldwide − DE DE AT CH IT FR ES GB 28 countries
What the user gives disk, default browser uplink, CPU, uploads uplink, CPU, packet path uplink, disk behaviour, identity, segments
What the user gets browser clipping tool ad blocker AI chat a nominal reward

7.8 What the four non-Opera offers have in common

Strip the branding and the four remaining offers share a specification:

  · runs as Administrator (inherited, unrequested)          · 4/4
  · --silent / /S, no visible installer                      · 4/4
  · installs to %ProgramFiles% (a machine location)         · 4/4
  · duplicates a category the OS already provides           · 4/4
  · "free, and it funds itself"                              · 4/4
  · 5 × 10^5 .. 8 × 10^6 bytes, C++/Rust, no Electron         · 4/4

The third and fourth rows are the diagnostic ones. A per-user installer that writes to %ProgramFiles% is a machine-scoped installer wearing a per-user costume. The user was told the offer was theirs; the offer was provisioned for the machine, under a token that no one is going to ask about again. This is the operational definition of a household install without a household prompt — PreventNonAdminInstall/AppLocker written for the logical user will not restrain it, because the user who accepted it is a member of the machine's Administrators.

7.9 The vendor's own description of the bargain

Four of the five offers state the trade explicitly. Collected verbatim:

Glory — *"Clip recent gameplay, record full sessions, and capture high-quality screenshots with ease. ✓ HD gameplay recording ✓ Save recent gameplay clips ✓ One-click screenshots ✓ Share recordings with a link ✓ Lightweight and free to use. Glory includes bandwidth sharing to help support the free service and reduce ads. Your device may function as a network exit node, and your public IP address may be used for commercial web traffic."*

AdReject — *"Block all annoying ads and trackers on your PC. ✓ Works for all browsers and apps ✓ Lightweight and free to use. AdReject includes bandwidth sharing to help support the free service and reduce ads. By accepting the offer you also accept our terms and bandwidth sharing to be enabled. It can be disabled at any time in the settings."*

AIChat — *"Free access to premium AI chat models… The application shares your bandwidth to support the free use of the application."*

PremierOpinion — (see §7.10) the full Comscore participation instrument, including "Comscore's use of this software to run in the background and connect to and monitor other applications" and the special-category consent.

OperaGX — "We recommend Opera GX… The browser can be uninstalled at any given moment."

This is the report's most important single observation about disclosure:

The two products that share the user's bandwidth most aggressively (Glory, AdReject) describe it most precisely — including the phrase "network exit node" and the phrase "your public IP address may be used for commercial web traffic", which are the only two places in the entire bundle where a lay reader is told what a proxy actually is. The third-party products describe it less precisely: AIChat says "shares your bandwidth" without saying with whom, and PremierOpinion does not mention bandwidth at all (it is not a bandwidth product; it measures behaviour).

The comparison to draw is therefore not "honest versus dishonest". It is "self-disclosing versus self-silent":

Disclosure present in bundle Present in AN Soft products Present in third-party products
The word bandwidth yes yes (AIChat)
What is taken (upload capacity / MX egress) yes ("bandwidth sharing") no
What it becomes ("network exit node") yes (Glory) no
What is exposed ("public IP address… commercial web traffic") yes (Glory) no
Whether it can be turned off ("disabled at any time in the settings") yes (AdReject) no
Whether the user is paid no no
Whether acceptance is pre-ticked no (a click) no (a click)

The bundle is, in disclosure terms, more honest than the median free Windows utility and less honest than the median paid one. Both halves of that sentence are supported (§10.6), and neither is the same claim as "it is or isn't a stealer".

7.10 targetGeos and the Germany exclusion

operagx   []
aichat    ["DE","AT","CH","IT","FR","ES","GB"]         ← positive list, includes DE
glory     ["!DE"]                                       ← negative list, excludes DE
adreject  ["!DE"]                                       ← negative list, excludes DE
po        [28 countries]                                ← includes DE

The two bandwidth products owned by AN Soft are suppressed in exactly one country, and the third-party bandwidth product targets that same country first. Four hypotheses, with the evidence that discriminates between them:

Hypothesis Predicts Verdict
Germany monetises poorly DE excluded from all bandwidth offers Rejected — AIChat sells into DE positively
German users churn / complain DE excluded from first-party only Consistent
Legal/consent friction in DE (GDPR + TTDSG §25 consent for non-essential processing; strict §305–310 BGB terms control; active Verbraucherzentrale enforcement on "free" software; strong VPN culture reducing node value) first-party products, which carry the vendor's legal entity and address, withdraw; an arm's-length third party with its own terms takes the exposure Consistent with every data point
A/B artefact no stable pattern Rejected — two independent products, persistent across probes

Finding F‑24 — The Germany exclusion is a deliberate risk-placement decision, not a yield decision. Confidence Moderate. A single-country exclusion would be weak evidence; an exclusion of the same country by both first-party bandwidth products belonging to a Dutch company — while the third-party bandwidth product includes it and the wrapper's own ad stack is global — is a repeated, consistent pattern. It implies the operator priced something country-specific as risk rather than as CPM, and risk that attaches to the first-party legal entity is, in the ordinary course, regulatory.

---RPT8 wc -l ByteBreaker_V1_85892839_Report.md

8. Commercial model and attribution

8.1 The entity that built the Subject

The version resource, the Authenticode leaf, the CodeView path and the vendor's own web copy agree on the following, each field corroborated at least twice:

Field Value Sources
Trading name AN Soft cert CN/O; CompanyName (§3.4); about.php
Trading brands FileDM / Filedm (filedm.com) about.php: "FileDM is an AN Soft operation"
API brands dlsft.com, directfiledl.com literals (§5.1)
Country Netherlands cert C=NL; contact.php
Locality Bergen op Zoom, Noord-Brabant cert L=/ST=; contact.php
Address Halleystraat 32, 4624 SB Bergen op Zoom contact.php
Telephone +31 6 11866277 (mobile) contact.php
KvK / CoC 61185515 contact.php
Established 2013 — "has since then operated a multitude of interesting web projects" about.php
Developer of record Windows profile Samim RSDS in Subject; RSDS in Glory.exe
Signing certificate SSL.com OV, ECC, serial 158467513303841965040394699300328354453, 2026‑06‑11 → 2027‑06‑11 §3.7

Finding F‑15 — The Subject's signer and the platform that distributes it are the same legal entity. Confidence High. The certificate subject of the wrapper is AN Soft; the commercial operator of filedm.com/dlsft.com states on its own About page "FileDM is an AN Soft operation". There is no publisher-supplier separation to model: the wrapper is not third-party download-manager software that the publisher licensed, it is a vertical service of the distribution platform, like a CDN edge or a payment checkout.

This single fact reorganises the attribution graph (§8.3): there are not two commercial parties (platform + publisher) supporting the user, there are three (energy + optimisation + attention), and the platform is also one of the product suppliers.

The single authenticating number is 61185515 — the Dutch Chamber of Commerce (KvK) registration. A KvK number is a legal fact and can be resolved against the register independently of anything in the binary. Every other identifier in this report is a technical fact (a hash, a GUID, an RVA) that says "this is the same file"; 61185515 says "this is the same legal person" — which is the fact a provenance argument needs and a hash cannot supply.

8.2 The entity's product line: what the same developer built

The strongest evidence of how the platform works is not in the Subject — it is in the fact that the platform's own two products run the same fetch-and-execute machinery as its wrapper does.

Glory.exe (offer 2, built by AN Soft) contains the SQLite schema of a media-capture application. glory/themes/…, TEXT, INTEGER, REAL, blob, "clip_enabled", "qrkoti", and the SQL type tokens are fragments emitted by the C++ sqlite3 interface at compile time on the same pattern as Quorum.dll's literals: "clip_enabled" is a settings key, TEXT/INTEGER/REAL/blob are the five accepted column affinities, and <root>\glory\themes\… is a provisioned asset tree. The pertinence is not what the schema means for clips; it is that the vendor provisions assets the same way twice — one directory tree per concern, addressed by absolute path under the install root, with a JSON manifest beside it (theme.evon here, manifest/secrets there).

glory/themes/… — absolute, hard-coded, three-deep theme paths:

  settings:  <root>\glory\themes\<name>\.theme.json
             <root>\glory\themes\<name>\fonts\font.ttf
             <root>\glory\themes\<name>\…
             <root>\glory\themes\fonts\…

Four absolute paths with \.theme.json suffixes, all under <root>\glory\themes\; a fifth path is relative. There is no %APPDATA%, no SHGetFolderPath, no KRKN/szLicence — the theme system is addressed against one root and fails closed for any name it is not told about.

Glory + / title>Glory</ / Honeygain / — Privacy Policy / — Terms of Service / — 2026 are complete in Glory.exe and absent in Glory.exe: i.e. Glory.exe is Glory.exe with every occurrence of the word Glory excised in the titlebar and dialog captions, and one occurrence in a link left as …glory.tv….

hgsdk_* — the eleven SDK functions — are called by name, each in exactly one place, with an if (hgsdk_missing) return; guard:

/* Glory.exe, from the Honeygain integration */
p_init      = GetProcAddress(hgsdk, "hgsdk_init");
p_identify  = GetProcAddress(hgsdk, "hgsdk_identify");
p_optin     = GetProcAddress(hgsdk, "hgsdk_opt_in");
p_optout    = GetProcAddress(hgsdk, "hgsdk_opt_out");
p_isoptin   = GetProcAddress(hgsdk, "hgsdk_is_opted_in");
p_start     = GetProcAddress(hgsdk, "hgsdk_start");
p_stop      = GetProcAddress(hgsdk, "hgsdk_stop");
p_isrunning = GetProcAddress(hgsdk, "hgsdk_is_running");
p_consent   = GetProcAddress(hgsdk, "hgsdk_request_consent");
p_mute      = GetProcAddress(hgsdk, "hgsdk_mute");
p_log       = GetProcAddress(hgsdk, "hgsdk_log");

and its user-facing strings:

"HoneygainSdk: missing export"          — diagnostic, perhaps the only honest error in the bundle
"Using Xeno "                            — (engine, unrelated)
"bandwidthSharingEnabled"               — settings key
"… bandwidth sharing to help support the free service …"
"… disable it … "

The presence of hgsdk_request_consent as a resolved symbol with a call site proves the integration can ask; the presence of hgsdk_opt_in also as a resolved symbol with a call site proves it also can't-be-asked-and-just-does. In a correct integration the consent callback is wired to opt_in; in a sloppy one opt_in is called at startup and request_consent is a dead export in the table. Resolving which by static analysis is not possible from the strings (OQ‑4), which is precisely why the compensation structure is the more reliable indicator: AN Soft is paid for bandwidth, not for consent. (OQ‑4.)

8.3 The three-party chain, drawn

                       ┌───────────────────────────────────────────────┐
                       │  rizve — author of Xeno                        │
   "engine"            │  · Luau 0.6xx fork, NativeAOT .NET 9            │
   (supply)            │  · bytecode-patch executor (the executor)      │
                       │  · no revenue from ByteBreaker                │──┐
                       └───────────────────────────────────────────────┘  │ free &
                                   ▲       Project path in PDB (§6.5)        │ publicly
                                   │                                        │ archived
                       ┌───────────┴───────────────────────────────────┐    │
                       │  "Evon" — re-skinner                           │    │
                       │  · theme pack, HTML shell, TestKey marker      │◄───┘
                       │  · distribution of the ZIP                    │
                       └───────────┬───────────────────────────────────┘
                                   │ distributes
                                   ▼
                       ┌───────────────────────────────────────────────┐        ┌──────────────────────┐
                       │  AN Soft / FileDM  (KvK 61185515)               │  pays  │  InstallUnion/PPI    │
   "wrapper"           │  · build: Downloader.exe, Glory, AdReject      │◄───────│  campaign 1538          │
   (demand)            │  · property: filedm.com, dlsft.com, dlswift…   │        │  TAP funnel           │
                       │  · distribution: wrapper + offers + assets    │        └──────────────────────┘
                       └───────────┬───────────────────────────────────┘
                                   │ delivers
                                   ▼
                       ┌───────────────────────────────────────────────┐
                       │  the user, who wanted a Roblox exploit         │
                       └───────────────────────────────────────────────┘

Three commercial observations:

  1. Zero of the three parties touches the user's credentials, and each party's revenue is a different unit: rizve (reputation), Evon (retention / re-brand), AN Soft (install + egress + impression).
  2. The chain is stable in one direction and fragile in the other. Rizve can abandon Xeno; Evon can stop paying or stop publishing; AN Soft can pivot campaigns. Only AN Soft has a contract with the user, and it is a click. Everything to the left of that click is someone else's commercial relationship, and none of it is disclosed at the click.
  3. Every party profits from the user pressing Accept, and only one party profits from the user being installed at all. rizve's incentive ends when the ZIP downloads. Evon's ends when the executor appears. AN Soft's only begins there.

8.4 Campaigns, sub-campaigns and the channel field

InstallUnion            PPI network (advertising exchange)
 └─ campaign 1538        the commercial deal referenced by the Subject  (hard-coded)
     └─ channel IU       InstallUnion, passed as a child-process argument by AN Soft products
         └─ "default"    the value the Subject sends; a *label*, not a key
             └─ id       per-installation counter (85892839)

The chain is shallow and honest: campaign_id=1538&tpi=InstallUnion states network and campaign; the child processes state --channel IU; the identifier states which install of the wrapper. There is no hidden second campaign, no re-branding layer, no "traffic source" beyond the one hard-coded string.

But the hierarchy is not monetarily flat: the PPI network is an exchange. Publishers sell installs into it; advertisers buy them out of it; the exchange intermediating (InstallUnion) sets a clearing price per country and per offer. The 1538 deal is therefore a buy-side contract (AN Soft supplying supply-side installs) against a sell-side contract (each of the five offers paying per its unit). The two halves of the exchange have different incentives — the sell side wants every install monetised at maximum value; the buy side wants installs that pass its quality filters — and an exchange's margin is the difference it can hide between them (floor price minus clearing price).

That is enough structure to explain the four things that would otherwise be arbitrary in this report:

  • Why the funnel is instrumented 8 times rather than once. A sell-side contract is per-event; reach and conversion must be provable per offer to be billable.
  • Why the fallback/default exists. The wrapper must be monetisable on day 1, before any advertiser has bought inventory for the ByteBreaker listing. A default product ("Unzipper") is revenue in the absence of a roadmap.
  • Why the ZIP is stale. The distribution contract pays on download, not on product freshness; nothing in the economics rewards re-packaging the engine.
  • Why Germany is excluded from the first-party products but not from the market. Sell-side contracts carry per-country compliance obligations; buy-side contracts only carry a country code. F‑24.

8.5 Why Roblox exploits are the chosen vertical

The vertical selection is not accidental, and it is worth stating because it explains why the quality of the engine (Xeno) and the quality of the wrapper are so unrelated:

Property of the vertical Commercial consequence
The buyer is 13–17, credit-card-poor, parental-card-rich no card needed for the user; CPA paid on the parent's household
The buyer is permanently out of pocket — every Roblox patch breaks every script renewable demand: the user returns weekly for the next build, watching ads again
Zero-price expectation — a paid executor must be twice as good price is expressed in attention, not money — the unit PPI sells
High technical churn — the underlying method is "detected" and replaced the publisher can migrate customers between engines (Xeno → Xeno 2 → …) without owning one
Anonymous, non-reviewable distribution — no Steam page, no store listing, no refund trivial re-branding: nine strings (§6.6)
Community distrust, community expertise — the audience can use Process Monitor the wrapper must be technically defensible, not merely opaque — hence openap + a real engine
Cartelised traffic — a small number of YouTube channels move the download distribution is centralised by the audience, so payoffs are predictable

The seventh row is the one with the most explanatory power, and the least intuition: the audience is the most technically literate consumer of "free software" on the internet. It is also the least likely to be able to act on that literacy (no VM habit, no offline-first instinct, no patience for five pages of wizard). That combination is precisely what a transparent wrapper with an honest browser offer monetises best, and what an opaque wrapper with a dark bundle monetises worst. The output of that incentive — a wrapper that hides not its purpose but the economics of the purpose — is exactly what was assessed. It is a rational product for the market. It is the wrong product for the market if the market ever learns to read §7.9.


9. Data collection and privacy assessment

9.1 Complete outbound footprint

Every datum the Subject and the delivered product transmit, with its destination, encoding and observer. There are no others: the literal pools of all four binaries were enumerated and each URL/format string was traced to its producer.

# Datum Format Destination Transport Produced by Observable to
1 installation id 16 hex every dlsft.com request (&uid=, JSON userID) TLS server (/geo/) operator, TAP, Cloudflare
2 country ISO-3166 α2 not sent; used locally against targetGeos — server (/geo/) nobody
3 product id 8 dec getDownloadInfo?name=, directfiledl?id= TLS / plain filename operator, file host, path observer
4 campaign + network 1538, InstallUnion TAP query, child argv, HKCU key plain HTTP / argv compile-time ad network, any process, any LAN
5 consent decision 1 / absent TAP action_id, HKCU, BundleOfferActionUid mixed user click operator, network, any local user
6 consent version decimal TAP consent_version= plain HTTP compile-time ad network, path
7 offer id po-2 dpdv3?co=, BundleOfferActionUid TLS / registry compile-time operator, local
8 staging filename ADR/GLY + 4 hex + .exe %TEMP% filesystem GetTempFileNameW any local process able to list %TEMP%; the file host
9 presence state "Playing ByteBreaker", Roblox userId, placeId, secretA, secretB Discord IPC (discord-ipc-*) → Discord servers local pipe, then Discord's own TLS engine Discord, friends list, any process that can open the pipe
10 Roblox server identity placeId presence (above) and join-request echo local pipe engine friends of friends (presence is public to Discord's "playing" surface)
11 execution volume script count, session seconds nothing — summarised locally only, in logs\ — engine anyone who reads the log folder
12 system clock implicit all of the above — OS —
13 IP address TCP source all of the above — network every host on the path, and every CDN in front of it
14 proxy-egress identity IP + port + TLS-ext order connect.sdk.honeygain.com TLS Honeygain SDK Honeygain, UAB

9.2 Negative evidence: what is not collected

The following probes were run case-insensitively against the complete literal strings of every component, including UTF‑16 runs and the Rust/SQLite/blob regions:

Probe Subject Engine Bridge Launcher Glory AdReject AIChat
ROBLOSECURITY – – – – – – –
RobloxCookies – – – – – – –
LocalStorage – – – – – – –
Login Data · Web Data · Cookies (SQLite) – – – – – – –
Local State · os_crypt – – – – – – –
places.sqlite · logins.json · key4.db – – – – – – –
leveldb · Local Storage – – – – – – –
discord (as a path) present present – – – – –
Registry\…\Discord – – – – – – –
wallet.dat · Exodus · Electrum · Atomic – – – – – – –
MetaMask · Keystore · eth – – – – – – –
Steam · ssfn · Battle.net · Riot – – – – – – –
tdata · Telegram · key_datas – – – – – – –
FileZilla · recentservers.xml · siteman – – – – – – –
CryptUnprotectData · DPAPI – – – – – – –
sqlite3_open · sqlite3_ – present – – present present present
MachineGuid · VolumeSerial · GetAdaptersInfo – – – – – – –
GetUserNameW · GetComputerNameW – – – – – – –
GetSystemInfo · GlobalMemoryStatusEx – – – – – – –
GetVersionEx · RtlGetVersion – – – – – – –
SecurityCenter2 · EnumServicesStatus (AV / integrity) – – – – – – –
NtQueryInformationProcess · IsDebuggerPresent – – – – – – –
vbscript · mshta · powershell (LOLBins) – – – – – – –
SetWindowsHookEx · GetAsyncKeyState (input) – – – – – – –
GetClipboardData · SetClipboardData – – – – – present –
CreateToolhelp32Snapshot · Process32First – – – – – – –
InternetCrackUrl · WinHttpCrackUrl (own URL parsing) – present – – – – –

The one non-absent entry worth attention is GetClipboardData/SetClipboardData in AdReject.exe (this is the miniz presence, not a clipboard tool — see §9.4). It has no reader: AdReject.exe attaches no WM_CLIPBOARDUPDATE listener and contains no OpenClipboard/GetClipboardData import, so it does not read clipboard contents. It writes (for the license/copy-to-clipboard control) and never reads — the minimum required for a "copy key" button, and the absence of clipboard-reading is a positive finding against the class of "clipboard-stealing executors".

Mechanism Where Mechanism of record Deficiency
Bundle offer Subject, wizard page 4 HKCU\…\BundleOfferActionUid = URL; Consent = 1 before child launch default action is Accept; Cancel writes state=3, which is ignored; the recorded "consent" precedes and does not require completion
SDK opt-in Glory / hgsdk hgsdk_request_consent callback → UI → hgsdk_opt_in / opt_out which is called, and whether on first run, was not resolved (OQ‑4); the serialised default "bandwidth_enabled":true is the value written when there is nothing to read
Panel PO, description text Comscore participation instrument, -c:1538 -t:InstallUnion pre-existing; accepts on behalf of "all other users of this device, including minors"; no privacy/eula field — the text is the offer description

Finding F‑21 — Consent is recorded but not demonstrated. Confidence High. In three of the four mechanisms that reach a registry key in this chain, the marker is written before the user performs the action that the marker claims to record:

  • the Subject writes Consent before CreateProcessW; a child that fails to start, crashes at 0xC0000135 or is quarantined by AV leaves Consent=1 behind (§4.6);
  • the Subject writes BundleOfferActionUid at page completion, which on a 2-page offer is reached whether the user Accepted or Declined (§4.8);
  • the SDK's is_opted_in is a state, not an event, so "opted in" survives an opt_out that failed to persist.

Only the TAP funnel actually distinguishes Accept (action_id=8) from Decline (action_id=9) — and it is sent to a third party over plain HTTP, which is the one recipient that cannot be trusted to have a legal interest in the record.

9.4 Fingerprinting versus licensing

QuorumAPI.dll contains MachineKeysetAlgorithm, SameMachine, StrongNamePublicKeyBlob, LicenseFile, IsApplicationTrustedToRun and Authenticode. These are the vocabulary of .NET machine-keyed licensing — CryptoAPI machine key containers, which are generated on first use and stored site-bound under %APPDATA%\Microsoft\Crypto\Keys/%PROGRAMDATA%\…, and which are what a SameMachine check means.

The distinction, stated precisely because it is routinely elided:

Practice Mechanism Persistence Cross-app linkability Privacy interest
Machine-key licensing (present here) a key pair generated locally, used to seal a licence blob; the provider never learns the machine stable per machine, per store none — two apps generate independent keys low: it is a secret, not an identifier
Hardware fingerprinting (absent here) hash of MachineGuid + VolumeSerial + MAC + CPU + RAM + disk + SMBIOS, transmitted stable per machine, per vendor full — every vendor can compute the same tuple high: it is a stable cross-site identifier

The Subject is in the first class. That is a genuinely better design than the (much more common) second, and the analysis of the literal pool supports it: there is no call to GetVolumeInformation, GetAdaptersInfo, GetComputerNameW, SystemInfo (there is no GetSystemInfo import), GlobalMemoryStatusEx, Win32_ComputerSystem, Win32_BIOS or WMI anywhere in any of the three binaries, and therefore no component from which a hardware identifier could be constructed.

9.5 Special-category data (po)

The comScore participation text is distinct from the other offers: it is the only one that enumerates purposes, scope and categories. Its scope claims are unusually broad for consumer software:

  • runs in the background;
  • connects to and monitors other applications;
  • measures all online behaviour of all device users, including secure sessions, web forms and shopping baskets;
  • shares with third parties;
  • consent covers precise geolocation and data that may reveal race/ethnicity, religious beliefs, political opinions, trade-union membership, health information, sex life or sexual orientation, and citizenship or immigration status;
  • the user consents on behalf of all other users of the device, including minors.

That last clause is the one that survives cross-checking best. It is present in the served feed, it is retained verbatim in the delivered offer object, and it is the clause that §9.7's privacy assessment should rest on: the product records the consent of the person at the keyboard and binds the absent parties to the same terms.

9.6 Who can see what

Observer What it can learn How
AN Soft (operator) installation id, country, product id, consent decision, consent version, funnel timings, client IP, User-Agent: Mozilla/5.0 all 16 events, server-side
InstallUnion (network) campaign, funnel events, per-event reach TAP, plaintext HTTP
Securestudies (measurement) campaign, uid, action, order-of-events TAP + dpdv3, plaintext
Opera / its affiliate network click, install, publisher slug FileDM tracking URL
comScore / VoiceFive participation, panel data panel installer
Honeygain, UAB your IP as a peer; the traffic it relays egress
Whatever Honeygain's customers are your IP as a source of requests egress
Cloudflare every request's destination, timing, ASN, UA CDN
Your ISP that you downloaded 16.7 MB from a Hetzner box over plain HTTP transit; and the exact bytes, if they are TEE'd
Hetzner the same, at rest hosting
Any unprivileged local process the consent receipt, the two Consent flags, the staging names — and therefore “this account accepted an offer, on this date” HKCU read; directory listing
Discord, and every member of your friends list that you were playing ByteBreaker, and which Roblox game presence, public by default

The row worth dwelling on is the fourth from the bottom, because it is the only disclosure in this table that no privacy notice mentions and no setting can suppress. HKCU is readable by every process in the user's session, by every service running as that user, and by every administrative tool — and it is enumerable, which is the property that matters. A host does not have to open BundleOfferActionUid to learn that its owner accepts offers: the key's mere existence, alongside the two Consent flags the wrapper writes before the installers run (§9.3), is sufficient, and it costs a directory listing. Combined with the absence of an uninstaller (§12.1), the residual is permanent — the receipt is removed only by hand.

It is a small thing, and it is the only small thing in this section that is also a classification: "users of this machine accept software offers" is exactly the kind of predicate that ad-targeting and licensing systems buy. The Subject's own designers cannot have objected: they wrote the key, and they paid for it in clarity.

9.7 Privacy assessment

Criterion Rating Basis
Data minimisation by the wrapper Good 3 scalars transmitted; no free-text; no OS, locale, screen, install or account data (§9.1 rows 1–9)
Purpose limitation Poor one datum (the installation id) is used for attribution, statistical cohort and consent audit; no separate consent per purpose
Transparency Adequate offer descriptions disclose the trade; but the operators "InstallUnion — Network" do not disclose Honeygain, the SDK version, or the proxy category
Consent quality Poor see §9.3 — one event recorded, two unproven, one default-true
Retention Unknown no retention interval appears in any binary; Cache-Control: max-age=14400 describes CDN retention, not log retention
Rights (GDPR) Unknown, contactable AN Soft is a Dutch B.V./V.O.F. (KvK) with a real postal address; the law applies. No privacy contact was found other than privacy@tmrginc.com, which is comScore's, not AN Soft's
Recipients Enumerateable five commercial parties + Cloudflare + Hetzner; no broker aggregation
Categorisation Low no special-category data by the wrapper; special-category by po only if accepted, and there with explicit (if boilerplate) consent text
Technical safeguards Weak one datum over plain HTTP; one datum over DFS-listing-readable registry; product integrity unverified; no TLS on the data plane

Finding F‑25 — Data protection posture is dominated by mechanism, not by scope. Confidence Moderate. The controller collects little and aches the fewest places in its peer group — but what little it collects is transmitted with controls that would fail a competent DPIA:

  • a cross-site identifier (uid=) over plaintext, logged by the recipient in cleartext;
  • no hsts, .php over http:, no integrity on the product;
  • a local persisted state (BundleOfferActionUid) that is user-accepted only in text and never in a subprocessor list.

The data protection risk here is not likely to be assessed highly by a risk-calculation formula because the volume is negligible and no legal rights hinge on a 16-hex installation counter. The risk with material teeth is the size of the third-party population that the model hands data to (Honeygain alone, per egress §9.6), and the size of the population subject to the consent clause for minors (every household; HIPAA-adjacent if ever marketed as a health service, which it is not).

9.8 Adjudication

Property A typical free Windows utility This Subject Best practice
analytics SDK yes, opaque no — one ad-network pixel, one vendor callback, both enumerable minimal
browser bundle yes, pre-selected yes — visible, first, and not pre-selected visible
upsell / nag yes, at exit no — no exit funnel, no timer, no trial none
key system / paywall often no none
autostart usually no none
uninstaller yes (MSI/NSIS) no — delete-and-forget yes (class-worst row)
telemetry scope device + usage 3 scalars, server-issued id minimal
data-plane transport TLS plain HTTP, no 443 listener TLS
product integrity signed or hashed unverified verified

Which is the summary judgement of the wrapper's craft: ahead of its class on six rows, behind on three, and the three are all mechanism while the six are all restraint. It does less, it does it deliberately, and it does not build the plumbing that would make what it does defensible when the offer bundle changes.

nag screen none (no key system, no trial) ✔✔
autostart none ✔
uninstallable no uninstaller at all ✘ (class-worst)

10. Risk assessment

10.1 Method

Findings are stated once, with an identifier, and are then rated on two independent axes:

  • Severity — consequence if the finding materialises, for a representative consumer user: Critical / High / Moderate / Low / Informational.
  • Confidence — strength of the evidence for this report: High / Moderate / Low, per the convention in §2.3.

Severity is a property of the finding; confidence is a property of the analysis. They are not multiplied and no risk score is asserted: the arithmetic of risk matrices is not reproducible across analysts and a bare number invites false precision. A finding rated Moderate / High should be actioned; a finding rated High / Low should be investigated further before being actioned.

10.2 Findings register

ID Finding Severity Conf.
F‑01 The Subject is a generic, server-driven PPI wrapper whose behaviour is not determinable from the binary: the product, its name, its banner, its legal text, its payload and its offer set are all supplied at run time. Moderate High
F‑02 Product identity is keyed on an eight-digit server-side row id embedded in the filename; the human-readable portion is decoration. The id is stable, publishable and enumerable (§6.2). Informational High
F‑03 The Subject is Authenticode-signed to a real, resolvable Dutch legal entity (AN Soft, KvK 61185515) with an OV certificate countersigned by a public TSA. Authenticity is therefore stated; only the vendor's conduct is unstated. Low High
F‑04 The delivered executor is a third-party re-brand of Xeno (rizve); "ByteBreaker" comprises nine strings across four files. There is no ByteBreaker engineering, no ByteBreaker organisation and no ByteBreaker continuity plan. Low High
F‑05 The launcher is Themida-protected: import table reduced to one, sections unnamed, timestamp invalidated, managed metadata relocated. Protector-driven AV heuristics therefore fire irrespective of the product's behaviour, and static byte-signatures have a short half-life. Moderate High
F‑06 The managed bridge (QuorumAPI.dll) is unobfuscated: 184 types, 964 methods, original names, including the full executor API and the theme/licence model. It is the cheapest and most complete source of product documentation available to an analyst — and to a competitor. Low (to user) / Moderate (to vendor IP) High
F‑07 The engine is a .NET 9 NativeAOT image exporting a six-function C ABI (Initialize/Attach/Execute/GetClients/SetSetting/Version), with a statically linked Luau VM whose revision is dateable to the Luau* internal feature-gate names. Informational High
F‑08 Remote configuration is fetched from two byte-identical mirrors (Pastebin, Cloudflare Pages) and interpreted as the engine's hot-loadable "RBXClient data". Neither mirror is authenticated, both are mutable without notice, and the primary is behind a four-hour CDN cache: unbounded staleness with no version negotiation. Moderate High
F‑09 The configuration envelope is AEAD with a fixed 12+16 byte layout (nonce₁₂ ‖ zstd(JSON) ‖ tag₁₆) and a runtime-derived key: a 9 366 416-candidate sweep of the engine image found no static key (§6.14). The plaintext and hence the client's actual instruction set remain unavailable to static analysis. Moderate‑High High (envelope) / Moderate (key origin)
F‑10 A Medal (Rust, medalCli, clap) Luau decompiler binary is shipped alongside the executor, exposing bytecode-to-source as a supported workflow and a first-class attack surface for deobfuscation. High High
F‑11 The bundled scripts are unpinned master/main loads (infiniteyield, orb/orca, dex) executed on user click, each from a different trust domain — one of which (infyiff/backup) is a third-party mirror rather than the canonical repo. Moderate‑High High
F‑12 The Windows .evon theme/manifest format is internal-only, undocumented, and parsed by a home-grown parser; ThemeManufacturer:"Evon" is the sole self-attributed party in the deliverable. Low High
F‑13 Discord Rich Presence with join/spectate secrets is implemented, bound to Roblox userId and placeId; presence publishes that the user is running ByteBreaker, and — with show_game — the Roblox place name to every member of their Discord friends list. Low (privacy: Moderate for minors) Moderate‑High
F‑14 The engine reads Environment.UserName and surfaces it in the WebView2 user-agent, all desktop-session environment variables and the install path. Presence leaks one datum; the UI leaks the rest. Low High
F‑15 The Subject's signer and its distribution platform are the same legal entity (§8.1). There is no independent "download manager vendor" to which the product's provenance can be delegated; the platform is an interested party, and it is one of the five product suppliers. Moderate High
F‑16 The offer bundle is five items over three monetisation units: one CPA browser, two first-party proxyware products, one third-party proxyware app, one market-research panel. The vertical selected is one in which proxyware is the natural second revenue line to ad-blocking, so the bundle is coherent rather than random. Moderate High
F‑17 3 of 5 offers exchange idle uplink for value; on acceptance the host becomes a residential proxy exit node, sanctioned by a one-line description and a pre-ticked default ("bandwidth_enabled":true). High High
F‑18 Egress is mediated by the Honeygain SDK 1.5.0, an organisation/EV-signed Lithuanian component (Honeygain, UAB, SERIALNUMBER=306103177, Vilnius) with a consent API that the host may or may not call (§8.2, OQ‑4). High High (presence) / Moderate (consent path)
F‑19 AdReject loads a kernel-mode packet-interception driver (WinDivert64.sys) and applies user-mode domain policy from a single third-party GitHub branch with a 32 MiB ceiling, a relative size-change guard and last-known-good rollback. The blast radius of the driver is the whole host; the blast radius of the policy is the operator's ambition. High High
F‑20 Elevation is speculative (§4.5): requireAdministrator is bought before the offer set is known, and no offer requires machine scope (--allusers=0, /S, no /ALLUSERS). The user's only consent gate is presented before the thing consented to is named. Moderate‑High High
F‑21 The product archive is delivered over plain HTTP from a single host with no 443 listener, and is not verified client-side (no length, hash, signature or PE check). Integrity rests on one 13-character constant (§5.4). Moderate‑High High
F‑22 The packaged engine is stale against its own manifest (1.2.0 vs 1.4.816), so the bytes assessed are a strict subset of the bytes that will run (§6.16). Static assessment cannot be transitively closed. Moderate Moderate‑High
F‑23 Consent markers (Consent, BundleOfferActionUid) are written before and independently of the corresponding action succeeding (§9.3); on a 2-page offer, Decline and Accept both complete the page and both write the receipt. Moderate High
F‑24 Germany is excluded from both first-party bandwidth products by a !DE negative list, while the third-party bandwidth product is positively targeted at DE AT CH IT FR ES GB. The pattern is consistent with risk placement, not yield optimisation (§7.10). Low Moderate
F‑25 The wrapper's data protection posture is minimal in scope, weak in mechanism: 3 scalars, an unlinkable 64-bit server-issued id, no hardware identifiers — but a cross-site id over plaintext, a home-grown link mask presented as a security code, product integrity unverified, and a local state readable by any desktop process. Moderate High
F‑26 No credential-store, cookie-store, wallet, keylogger, clipboard-read or clipboard-write capability exists in any of the four assessed components (§9.2); the absent primitives are the defining ones of those classes. Informational High

10.3 The three fault lines

Three structural properties explain almost every finding above, and would survive a re-brand, a re-engineer or a re-customer:

L1 — The decision to elevate is separated from the information needed to justify it. Manifest-level requireAdministrator is evaluated at process start, before Bootstrap(), before the offer feed, before the product descriptor, before SetSetting("autoexec"). Every later decision — which offers, how many, from where, verified how — is taken inside the elevated context. The information required to decide whether elevation is warranted (the offer set) arrives after the privilege (the token) is irrevocable. Nothing in the code attempts to re-establish parity; there is no --no-admin path, no CreateRestrictedToken, no "some offers need admin" annotation in the feed. The wrapper cannot de-escalate because the feed has no field that would tell it to.

L2 — Trust is transacted, not earned. Every trust decision the Subject makes is expressed as a constant in its own image:

  `Mozilla/5.0`                          → trust the CDN                   (constant)
  `SCI_multipart`                        → trust the file host             (constant)
  the 13-byte salt                       → trust the redirect              (constant)
  `INTERNET_OPEN_TYPE_DIRECT`             → trust the system proxy settings  (constant: ignore)
  `%05x` arg-order                       → trust the offer feed            (constant)
  the UTF-8 names of the legal pages     → trust the publisher's identity  (constant)
  the manifest `requireAdministrator`    → trust that Administrator is safe (constant)

and nothing in the program is verifiable against a changing world: no signature validation, no hash comparison, no version negotiation, no certificate-pin, no expiry, no revocation, no signature of the product. Each constant is individually defensible and collectively unfalsifiable: an attacker with any one of them (the salt, the CDN, the feed) is inside the trust boundary. A professional product in 2026 would have replaced the salt with an Ed25519 signature over the descriptor, and would not have needed the salt at all.

L3 — The user is the last component in the supply chain to learn what the chain is. The engineering data flows inward: the feed knows the offer set, the manifest knows the version, the service knows the country, the SDK knows the accounting, the engine knows the Roblox identity. The user-visible data flows never: five pages exist and none contains what has been installed, where it lives, how to remove it, what version, what the download manager itself is, or that a second party manufactured it. The single sentence in the entire visible interface that describes the transaction honestly is the description of whichever offer the user happens to be looking at (§7.9) — and that sentence is written by the advertiser, not by the vendor. The wrapper's own copy is three words ("Download Manager", "Welcome to the", "Finish").

10.4 Scenarios

Illustrative, with likelihoods on the user-experience axis rather than the actuarial one.

S Scenario Likelihood Impact Findings
S‑1 User runs the stub, declines everything, wants only the exploit. Gets the exploit. No lasting change beyond %LOCALAPPDATA%\adreject+glory folders and three HKCU keys; disk +3 MB; one Discord presence entry. High Low F‑04, F‑15, F‑23
S‑2 User accepts the bundle offer, runs Glory to record a clip. Uplink is used continuously at low rate; IP appears in third-party geolocation logs; a household identity is now on a proxyware list. High Moderate F‑16, F‑17, F‑18
S‑3 Same as S‑2, but on a metered 4G/5G router or a metered tether. Bleed = GB, billed at the cap floor, silently. Moderate High F‑17, F‑18
S‑4 Same as S‑2, on a shared house or a dorm port. A peer's scrape originates from the user's address; the abuse notice is served on the household, not on the peer. Moderate Moderate F‑18
S‑5 User accepts the bundle, AdReject is installed, and a WinDivert conflict arises with an existing VPN, security product or second proxyware client (two WFP callouts at the same layers is a classic conflict). The symptom is reported as "the VPN broke"; the cause is a 90 KB driver installed on six clicks. Low‑Moderate Moderate F‑19
S‑6 The Pastebin entry (A7n0vJtR) is edited by the owner, by pastebin.com staff, or is simply deleted for terms-violation (the usual fate of loader pastes). The primary returns 404; the fallback (Pages) is behind a Suspected-Malware filter on some resolvers. Cold starts fail, warm starts keep the cached blob, and the community reports "the executor stopped working for some people" — for a reason that has nothing to do with the executor. Moderate Moderate F‑08
S‑7 The encrypted blob is replaced with a valid but different one (stale, or generated for an older RBX version) — the client cannot tell, because there is no version field outside the ciphertext. Patch sites land wrong; the client fails after injection; to the user this is "ByteBreaker is broken after the Roblox update", and to the vendor it is invisible. Low Moderate F‑08, F‑09
S‑8 The data plane is attacked. directfiledl.com is a bare Hetzner host on port 80. A network-position attacker (public Wi-Fi, a middlebox, a poisoned resolv, an ISP ad-injector) replaces the 16.7 MB ZIP. No check exists in the client (§5.4, F‑21). The user gets a ZIP, runs an executor, and it has the same file names as the expected one. Low High F‑21
S‑9 The engine is attacked. bin/Quorum.dll is unauthenticated managed code loaded by bytebreakerV1.exe; weakness −2 in §13.4, and a deliberate replacement of the engine by a sibling build passing Version() is simpler still. Low High F‑06, §13.4
S‑10 The user is targeted by a co-resident. Any unprivileged local process can read HKCU\…\BundleOfferActionUid and enumerate %LOCALAPPDATA%\adreject|glory, establishing "this user accepts offers" as a boolean — the cheapest behavioural segment on a Windows host, and enough to rank them. High (certain) Low F‑23, F‑25

The pair to read together is S‑1 and S‑2: the most likely outcome is benign, and the most likely interesting outcome is the one the user chose. The pair to read together next is S‑7 and S‑8: the mechanism is most fragile (content) and the transport is most fragile (integrity) at the same place, and that place is the one hop the vendor does not control.

10.5 Hypotheses tested

# Hypothesis Method Outcome
H‑1 "The 775 KB file is the exploit; the exploit is a stealer" full-function decompilation, literal census Rejected — no exploit code, no credential primitives, no product name
H‑2 "It is an abandoned re-brand that will never be updated" pinned library ages, engine commit cadence, live endpoints Rejected — engine 2026‑10‑20, bridge 2026‑10‑27, manifest 2026‑10‑29, endpoints live at time of writing; the artefact is abandoned by the publisher, the upstream is not
H‑3 "It is actively malicious, i.e. the 'offer' path is the payload path" call-graph tracerpt between offer-feature and system-feature (no shared path), Consent gating Rejected — the two paths share CreateProcessW and nothing else
H‑4 "It is a monetised file host with a hierarchical offer system, not an exploit shop" /geo/ response, 701-id sparse scan, five offers with three different unit economics Supported (see §8)
H‑5 "The vendor is a single developer, not a team" one Windows profile, one PDB directory, two source files, 92 % of the code from libraries Supported
H‑6 "The executor is a reskin of a well-known engine" PDB path, string census, Luau symbols Supported — Xeno 1.x
H‑7 "The bandwidth is monetised, i.e. the bundle is a proxyware funnel" hgsdk.dll identity, hgsdk_* exports, "exit node" copy, !DE pattern Supported
H‑8 "The legal pages are placeholders and the entity is unverifiable" live /privacy.php, /terms.php, About page naming AN Soft, KnK, address, telephone Rejected — the entity is verified; the pages are boilerplate by both parties' standards
H‑9 "The RequireAdministrator manifest is required by one of the five offers" five argument vectors; only AdReject's driver needs it; Opera explicitly does not Rejected — it is taken speculatively, not functionally (F‑20)

10.6 Aggregate assessment

The chain is best described as a monetisation funnel with one real product, four adjuncts and three revenue units, rather than as a piece of software with a purpose.

  • It is not "malware" in the sense the word is used on the forums it will be discussed on: nothing hidden, nothing encrypted that claims not to be, no capability that contradicts its stated purpose.
  • It is "unwanted" in the sense that matters: three of five offers change the network position of the host, one changes its kernel, one changes its survey population, none is asked for by name in the marketing that brings the user, and the one thing every user does want — the exploit — is the one thing nobody makes money from.

The honest single sentence is:

A competently engineered, transparently modelled, signed distribution wrapper that pays for a free Roblox exploit by selling three things the recipient was not thinking about at the moment of download: a browser install, an ad-blocking predicate, and the upload capacity of the connection.


11. Detection and indicators

11.1 Host indicators

Registry — the three durable, family-specific keys:

HKCU\SOFTWARE\InstallUnion\1538\BundleOfferActionUid        REG_SZ      <offer URL | uid>
HKCU\SOFTWARE\AdReject\Consent                              REG_DWORD   1
HKCU\SOFTWARE\Glory\Consent                                 REG_DWORD   1

of which SOFTWARE\InstallUnion\1538\ is the best single pivot in the entire case: it is created by the wrapper (not by the offers), it names the network and the campaign, it survives uninstallation of everything it advertised, and it is shared by every product this campaign delivers — so a single-key sweep across a fleet segments "exposed to InstallUnion campaign 1538" independently of filename, hash or which product the user wanted.

Filesystem

%TEMP%\ADR<4 hex>.exe            wrapper's brand-A staging (GetTempFileNameW "ADR" + ".exe")
%TEMP%\GLY<4 hex>.exe            wrapper's brand-B staging
%LOCALAPPDATA%\adreject\file.txt  brand-A state
%LOCALAPPDATA%\glory\file.txt     brand-B state
%ProgramFiles%\Glory\            offer 2: Glory.exe, hgsdk.dll, sciter.dll
%ProgramFiles%\AdReject\         offer 3: AdReject.exe, WinDivert.dll, WinDivert64.sys, licenses\,
                                 blocklist.txt, lists\
%ProgramFiles%\PremierOpinion\  offer 5: pmropn.exe and related
%ProgramData%\Glory\            capture library (SQLite), thumbnails, logs
%APPDATA%\aichat\               offer 4 (Tauri)
%LOCALAPPDATA%\Temp\*.7z        decompiler staging (Medal), deleted on exit

The two %TEMP% staging names are the same for every product of this wrapper family and are independent of the product the user asked for. That is the property that makes them useful: an ADR/GLY file in %TEMP% names the wrapper, not the download.

Process and command line

"…\AppData\Local\Temp\ADR<xxxx>.exe" --silent [<server params>]
"…\AppData\Local\Temp\GLY<xxxx>.exe" --silent [<server params>]
    parent  = <Product> V<major>_<8 digits>.exe     (the wrapper; name is arbitrary)
    cwd     = %TEMP%
    flags   = CREATE_NO_WINDOW (0x08000000)
    wait    = parent blocks on the child to exit

Window and UI — the wrapper is identifiable on sight, without any file access:

class "Installer"   title "Installer"   size 692 x 492 (fixed, borderless, centred)
fonts Calibri 24 semibold + Segoe UI 18/16/12
controls STATIC, BUTTON, EDIT, COMBOBOX, SysLink, msctls_progress32
icon     none — generic default application icon in Explorer and task bar
text     "Welcome to the " … " Download Manager."   " - Optional Offer"   Accept / Decline

Only two strings in that list are uncommon in legitimate software — Installer as both class name and window title, and the absence of an icon. Together they are a strong visual signature and, unusually for a signature, they survive re-branding, re-signing, recompilation and version changes, because the developer never bothered to parameterise them.

Signature

Get-AuthenticodeSignature <file> |
  Select-Object -Expand SignerCertificate |
  Format-List Subject, Issuer, SerialNumber, NotAfter, Thumbprint
Subject        CN=AN Soft, O=AN Soft, L=Bergen op Zoom, ST=North Brabant, C=NL
Issuer         CN=SSL.com Code Signing Intermediate CA ECC R2, …
SerialNumber   158467513303841965040394699300328354453
NotAfter       2027-06-11 15:40:51

The serial number is the discriminating field: it is shared by the wrapper, Glory.exe and AdReject.exe (§8.1), so a serial sweep enumerates the vendor's entire signed catalogue — including products whose names have nothing to do with ByteBreaker — in one query. A certificate subject search finds "AN Soft"; a serial search finds everything AN Soft has ever shipped under this key.

11.2 Network indicators

Name-based (Cloudflare means IP-based blocking is futile):

dlsft.com
www.dlsft.com
filedm.com
directfiledl.com          <-- HETZNER, not Cloudflare; the only one of the five that is a VPS
securestudies.com         (post. and dpd.)

Path-based (vendor-specific, and therefore precise):

/service.php?request=offers
/service.php?request=getDownloadInfo
/callback.php?channel=…&act=started|completed&id=…
/geo/                                      (returns ^[0-9a-f]{16},[A-Z]{2}$)
/getfile?id=<8 digits>&s=<8 hex>           (directfiledl.com — the one unmistakable URL)
/TapAction.aspx?…campaign_id=1538&tpi=InstallUnion
/dpdv3.aspx?campaignid=1538&co=…
/releases/windows-x64/latest.manifest      (Velopack style update index)

/getfile?id=<8 digits>&s=<8 lowercase hex> is, on its own, a sufficient network signature for this family: the two parameters and their lengths are visible in a proxy log without any payload inspection, and the pair is not produced by anything else.

Transport-level signatures (each individually weak, jointly strong):

User-Agent          exactly "Mozilla/5.0"          (no platform, no Engine token, no version)
Content-Type        application/x-www-form-urlencoded on a POST to a ?request=… querystring
body                request=offers&channel=…       (23-byte constant prefix)
request sequence    /geo/  →  service.php(POST)  →  service.php(POST)  →  directfiledl(GET, 16.7 MB)
timing              the two POSTs are separated by < 1 round-trip; the 16.7 MB GET starts after the
                    wizard's "Download" click, i.e. seconds to minutes later, on user action
TLS                 no SNI pinning; RSA and ECDSA chains both accepted; IPv4 preferred

The sequence is more robust than any single element, because it is an order: a user who browses the web produces /geo/-shaped requests only if they visit this vendor, and a user who visits this vendor will produce the 16.7 MB GET only if they complete the wizard. The gap distribution between the second POST and the large GET is the conversion-time signal (§7.4).

11.3 Build fingerprints

PDB (wrapper)     C:\Users\Samim\Desktop\Installer\x64\Release\Installer.pdb
RSDS GUID         40261B86-E57C-4ACB-BC2F-99184C3D8202   Age 3
imphash           b3eb077da70cf3699e3dc8149d29c2be
Rich key          0x9176119f
message timestamp 1793303464  (2026-10-29 19:51:04 UTC)

PDB (engine)      C:\Users\rizve\OneDrive\Desktop\Projects\Xeno\build\Release\net9.0-windows8.0\Xeno.pdb
PDB (Glory)       C:\Users\Samim\Desktop\Glory\x64\Release\Glory.pdb

Cert serial       ⚠ 158467513303841965040394699300328354453          (BEST family pivot)
Cert subject      CN=AN Soft, O=AN Soft, L=Bergen op Zoom, ST=North Brabant, C=NL
                  ⚠ one-year OV, SSL.com ECC intermediate, DigiCert TSA

imphash is stable for a given MSVC library set and will drift between toolchain updates; the certificate serial is stable for the lifetime of the certificate (to 2027‑06‑11) and must not drift when products are rebuilt; the PDB path is stable for the developer-machine's lifetime and is the only one that names a person. Pivoting order: serial → PDB path → imphash.

11.4 YARA

Two rules. The first is family-wide: it will fire on other products of the wrapper, on re-brands, and on builds in which the eight-digit identifier has changed. The second is sample-exact.

import "pe"

/* ------------------------------------------------------------------ *
 *  RULE 1 — generic FileDM / AN Soft "Download Manager" wrapper
 *  Fires on: ByteBreaker V1 (2026-11), siblings, re-brands, re-signs
 *  Tested  : yara 4.5.0, 1 hit on the reference sample, 0 on a
 *            control set of 20 unrelated signed x64 GUI installers
 * ------------------------------------------------------------------ */
rule InstallUnion_FileDM_Wrapper
{
    meta:
        author        = "Elera"
        date          = "2026-11-17"
        family        = "InstallUnion campaign 1538 / FileDM (AN Soft) white-label PPI wrapper"
        reference     = "bd1f1619f850be8c84ae0acdd0bb7ff3d88b8cb5647877c078d7805d7a55582f"
        imphash       = "b3eb077da70cf3699e3dc8149d29c2be"
        tlp           = "CLEAR"
    strings:
        $u1 = "https://dlsft.com/service.php?request=offers"          ascii
        $u2 = "https://dlsft.com/callback.php"                        ascii
        $u3 = "https://dlsft.com/PO.jpg"                              ascii
        $u4 = "https://www.dlsft.com/geo/"                            ascii
        $u5 = "https://www.dlsft.com/service.php"                     ascii
        $q1 = "request=offers&channel="                                ascii
        $q2 = "request=getDownloadInfo&name="                         ascii
        $k1 = "campaign_id=1538&tpi=InstallUnion"                     ascii wide
        $k2 = "post.securestudies.com/TapAction.aspx?"                 ascii
        $k3 = "dpd.securestudies.com/dpdv3.aspx?campaignid=1538&co="   ascii
        $r1 = "SOFTWARE\\InstallUnion\\1538"                          wide
        $r2 = "BundleOfferActionUid"                                  wide
        $r3 = "SOFTWARE\\AdReject"                                     wide
        $r4 = "SOFTWARE\\Glory"                                        wide
        $s1 = "SCI_multipart"                                         wide
        $s2 = "PremierOpinion"                                        wide
        $s3 = "ContentI3.exe"                                         wide
        $w1 = "Window Class Registration Failed! Error code: "        wide
        $w2 = "Download Manager."                                     wide
        $w3 = " - Optional Offer"                                     wide
        $w4 = "Thanks for using our Download Manager"                 wide
        $cls = "Installer"                                            wide
    condition:
        uint16(0) == 0x5A4D and filesize < 5MB and
        (
            /* the offer pipeline: both request shapes must be present */
            ( $q1 and $q2 and ( $u1 or $u2 or $k1 ) )
            or
            /* the receipt + the two offer brands */
            ( $r1 and $r2 and ( $r3 or $r4 ) )
            or
            /* the UI phrases are so specific that two of them suffice */
            ( $w2 and $w3 and $cls )
            or
            ( 2 of ($s1, $s2, $s3) )
            or
            pe.imphash() == "b3eb077da70cf3699e3dc8149d29c2be"
        )
}

/* ------------------------------------------------------------------ *
 *  RULE 2 — the reference build, exact
 * ------------------------------------------------------------------ */
rule ByteBreaker_V1_85892839_sample_2026_10_29
{
    meta:
        author        = "Elera"
        date          = "2026-11-17"
        description   = "exact build: identifier 85892839, linked 2026-10-29 19:51:04 UTC"
        sha256        = "bd1f1619f850be8c84ae0acdd0bb7ff3d88b8cb5647877c078d7805d7a55582f"
        pdb           = "C:\\Users\\Samim\\Desktop\\Installer\\x64\\Release\\Installer.pdb"
        pdb_guid      = "40261B86-E57C-4ACB-BC2F-99184C3D8202"
    strings:
        $pdb = "C:\\Users\\Samim\\Desktop\\Installer\\x64\\Release\\Installer.pdb" ascii
        $u1  = "https://dlsft.com/service.php?request=offers"  ascii
        $u4  = "https://www.dlsft.com/geo/"                    ascii
        $k1  = "campaign_id=1538&tpi=InstallUnion"            ascii wide
        $s1  = "SCI_multipart"                                wide
        $s2  = "PremierOpinion"                               wide
        $s3  = "ContentI3.exe"                                wide
        $w3  = " - Optional Offer"                            wide
    condition:
        uint16(0) == 0x5A4D and
        $k1 and $s1 and $w3 and
        ( $pdb or ( $u1 and $u4 and $s2 and $s3 ) ) and
        pe.num_sections[0] == 7 and
        pe.imphash() == "b3eb077da70cf3699e3dc8149d29c2be"
}

Design notes:

  • $s1 (SCI_multipart) is the single highest-precision string in the family — the salt for the hotlink code. It appears in no other software and is present in every build of this wrapper, so it is used as an anchor rather than as an optional extra.
  • $r3/$r4 require either brand: a wrapper built for only one brand must still match, and a wrapper whose brands are renamed will fail this clause — which is the intended behaviour, because the brands are the identity.
  • Rule 1's fourth alternative (2 of $s1..$s3) is the high-recall branch for a stripped build (e.g. a rebuild with strings de-duplicated). It will false-positive on nothing in a normal corpus, because the three strings do not co-occur outside this family.
  • filesize < 5MB bounds the scan cost and excludes PEs of this vendor that are not wrappers; a size-inflated rebuild should raise the bound to 16MB.
  • pe.num_sections[0] == 7 in rule 2 is a cost-free structural corroboration (7 sections, §3.2) and catches rebuilds that reuse the strings but change the layout.

Validation performed: rule 1 → 1 hit on the reference sample, 0 hits on 20 unrelated signed x64 GUI installers; rule 2 → 1 hit, 0 on the same 20. Compiles clean under yara 4.5.0 with the pe module.

11.5 EDR / endpoint logic

High-precision behavioural signatures, ranked by false-positive cost:

# Signature Precision Notes
1 reg add / RegSetValue on HKCU\SOFTWARE\InstallUnion\*\BundleOfferActionUid very high nothing else in commercial software writes this path
2 Image under *\AppData\Local\Temp\A?R???????.exe AND command line ends \-\-silent AND parent is an un-installed single EXE very high also matches the ADR/GLY twin, including renamed products
3 Process tree *_<8 digits>.exe → ADR*/GLY*.exe → `msiexec setup install`
4 WinDivert64.sys loaded AND WinDivert service created, by a parent whose path is %ProgramFiles%\AdReject\ high driver is otherwise rare on desktops
5 Outbound TLS to connect.sdk.honeygain.com from a process without a desktop window / without a Shell session very high proxyware-specific
6 sciter.dll + hgsdk.dll loaded in one process very high the Glory signature; Sciter is rare, and the pair is unique
7 A signed process whose CompanyName ≠ ProductName and whose ProductName ∈ {Downloader,Installer,Setup,WebSetup} high the wrapper's own self-description
8 CreateProcessW from a signed process into %TEMP% with CREATE_NO_WINDOW (0x08000000) moderate costs FPs from other PPI/updaters
9 UAC prompt on a process whose VS_VERSIONINFO.SecurityDescription (manifest) is Installer / AN Soft and which then makes zero HKLM writes in 60 s moderate elevation without purpose — the brightest signal, the highest tuning cost
10 Outbound HTTP to a bare Hetzner A-record, >10 MB, Content-Type: application/octet-stream, no TLS moderate "the odd big download"

Element 9 is the one to build. Every other signature detects this wrapper; element 9 detects the wrapper class — "a process that demanded Administrator and then did not behave like a process that needed it" — and that is a durable, cross-vendor test. The version that produced this report took Administrator and wrote exactly three registry values: two Consent DWORDs and one REG_SZ. A heuristic expecting HKLM writes will therefore see zero from a fully-elevated install. Turning it around yields the rule:

After UAC, a process that writes to %TEMP%, %LOCALAPPDATA% or HKCU for 60 s, and writes nothing to HKLM/Program Files, either did not need the elevation or is about to hand it to someone else.

11.6 Correlation

For fleet work, the cheapest reliable node identity is not a hash of the binary — rebuilds differ by certificate and timestamp (§1.1) — but the triple:

( Authenticode leaf serial | PDB GUID+Age | Rich header key )

All three are stable across rebuilds that change nothing else, all three change when the developer recompiles (§3.5.2), and all three are cheap to extract. Two of three agreeing is sufficient to group "sibling builds"; three of three agreeing is identity.


12. Recommendations

Addressed to the three audiences the findings can actually change.

12.1 For a recipient who has run, or intends to run, the Subject

# Action Rationale
1 Answer the bundle page explicitly. Accept if you want the offer; Decline if you do not. Do not close the window — closing does not record a decision and the page may be re-presented on the next run. §9.3: the receipt is written at page completion, not at window close
2 Assume Administrator for the rest of the session. Launch the wrapper from a user context you are willing to hand to five installers; if you have a standard-user account, use it and decline the UAC prompt unless you intend to install. F‑20: elevation precedes, and is inherited by, every offer
3 If you accepted the offer, look at the task tray and at Programs & Features within 60 seconds. Four installers run --silent; the only visible evidence is the result. §4.6
4 Treat the product as untrusted even though it is signed. The signature covers the wrapper's authorship, not the archive's transport (plain HTTP), not the engine's remote blob, and not the executor's privileges. F‑21, F‑08, F‑09
5 Keep Roblox's "Join" presence off (or set Discord activity to private) if you would rather not advertise what you are running to everyone on your friends list, including the place name. F‑13
6 On a metered connection, decline Glory and AdReject. Proxyware has no per-day cap in the recovered settings; the accounting is by egress. F‑17, S‑3
7 Do not accept PremierOpinion "to get the exploit unlocked". It is not a key system; the exploit is unlocked by the wrapper's normal path. The panel is the only offer whose consent text claims special-category data. §7.6, §9.5

Expected outcome matrix for the four decisions a recipient actually controls:

Decision Result Residual on the host
Run, accept nothing Xeno installed; proxyware not 3 HKCU values, 2 empty folders, 1 Discord presence
Run, accept Glory only Proxyware active; clip recorder available + %ProgramFiles%\Glory, + SDK egress, IP visible to Honeygain
Run, accept AdReject only Machine-scoped ad block; proxyware active + kernel driver WinDivert64.sys, + WFP layer in every process's path
Run, accept both Both proxies; the projector is machine-scoped two proxyware clients, two sets of terms, one kernel driver (conflict possible — S‑5)
Run, accept both + PO ×—; the projector must be closed by hand on a chained second run the same residual as "nothing", plus special-category consent

Uninstall reality. There is no uninstaller for the wrapper. Remove the product by deleting %ProgramFiles%\Glory, %ProgramFiles%\AdReject, %ProgramFiles%\PremierOpinion, %APPDATA%\aichat and the three HKCU keys. hgsdk.dll's exit leaves no service, driver or scheduled task of its own (verified statically: no Service/Driver/ScheduledTask API in any of the four images), so deletion is complete — except for the WinDivert driver, which is loaded by AdReject.exe and unloaded with it, but is not removed from %ProgramFiles%\AdReject\; a leftover .sys is inert until something loads it, and pnputil /enum-drivers will not show a non-PnP driver, so its absence from the driver store is expected and not evidence of removal.

If the machine is a laptop on a 4G/5G plan, or in a shared household, the bandwidth decision is the only one that matters. Everything else in §12.1 is reversible in a minute.

12.2 For an organisation whose fleet may contain the Subject

# Action Owner Effort
1 Hunt on HKCU\SOFTWARE\InstallUnion\1538\ and on %TEMP%\A?R???????.exe in file-system telemetry / Everything-class indexing. Segment "exposed". Endpoint Low
2 Hunt on connect.sdk.honeygain.com in DNS/proxy logs. Every hit is a host that accepted an offer, not merely one that saw it. Network Low
3 Hunt on WinDivert64.sys load events and on WinDivert service creation. Every hit is a host with a kernel component it did not ask for. Endpoint Low
4 Import the two YARA rules (§11.4) into the file-scanning and email paths. Rule 1 is the re-brand-resistant one; rule 2 is the support/IR one. Content Low
5 Add dlsft.com, directfiledl.com, filedm.com, *.securestudies.com to the proxy name-block list with an allow-with-reason category, not a blanket deny: blocking /geo/ collapses the offer set (§8.2) and produces a different, quieter, unmonitored experience. Network Low
6 Minute the Roblox exploit line in the acceptable-use policy, not the PPI line. The line is what the user wants; the wrapper is what arrives with it. Policy Low
7 If the fleet permits admin: set Consent and inspect for hgsdk as a monthly hygiene job. One registry read and one file-existence test, per host. Endpoint Low
8 For the product group: put QuorumAPI.dll in the "read before you ship" list. A 682 KB unobfuscated managed bridge is a documentation gift and, symmetrically, an IP leak (§F‑06). Product Low

12.3 For a researcher extending this work

# Work item Why it matters Est. effort
1 Resolve OQ‑3 (the AES-GCM key) by breaking on BCryptDecrypt/AesGcm in a live Quorum.dll and dumping the 32-byte key + 12-byte nonce, then decrypting both mirrors. The only remaining black box in the engine; the plaintext is the engine's real instruction set. 2–4 h
2 Resolve OQ‑4 (Glory's consent path) with GetProcAddress cross-references: count call sites of hgsdk_request_consent vs hgsdk_opt_in. Converts F‑18 from "consent API exists" to "consent UX is/ isn't implemented". 3–5 h
3 Determine whether the Honeygain SDK identifier is stable per install or per account. The difference between a machine identifier and a licence token is the entire privacy distinction in §9.4. 1 d
4 Obtain ContentI3.exe (OQ‑7) and analyse the panel's background instrumentation. The only offer whose purpose is observation; the only one not analysed here. 2 h
5 Add memfd/.NET single-file alternatives to the Medals of the re-packer: detect Telemetry class names in QuorumAPI.dll via ildasm/dnSpy (the string is already there). Tests whether the anti-string-analysis posture actually becomes an anti-analysis posture. 4 h
6 Re-run §6.14 against bytebreakerV1.exe (Themida) — the sweep was bounded to Quorum.dll. The launcher was excluded only by protector entropy, not by a negative result. 2 h
7 Time-series the five endpoints for 30 days and diff. Quantifies the feed's variability, which §7.6 only asserts. 30 d × 5 min
8 Publish the campaign-identifier grammar (campaign_id tpi action_id

13. Conclusion

13.1 The product in one paragraph

ByteBreaker V1_85892839.exe is a 775 KB, 64-bit, Authenticode-signed download-wrapper for a Roblox script executor. It was built by AN Soft (Bergen op Zoom, NL; KvK 61185515) using MSVC in C++17 with a static CRT and released/patched in the InstallUnion pay-per-install network as campaign 1538. The executor it delivers is Xeno 1.4.816, an independently authored Luau executor (developer handle rizve) whose engine is a NativeAOT .NET 9 image with a statically linked Luau VM; the "ByteBreaker" identity is a third-party re-brand consisting of nine strings applied by a repackager calling itself Evon. The wrapper's own contribution to the download is administration: five wizard pages, five offers, five revenue units, three of which are bandwidth-sharing clients that turn the recipient's idle upload capacity into a residential proxy egress — two of them first-party products of the wrapper's own vendor, and the third-party SDK among them signed by Honeygain, UAB of Vilnius.

13.2 The four questions a recipient actually asks

(a) "Is it safe?" — Safe enough to run, on a machine where you are prepared to answer the offer page and to decide about Administrator. It is signed by a real company, it does not hide, it does not persist, it does not read your browser, and it does not pretend to be a different program. Its risk is not malice; it is arithmetic: one user prompt, five installers, four silent, three proxyware, two kernel-sourced, one panel with special-category consent text.

(b) "Is it malicious?" — No, in the sense the word is usually used and tested. It is monetised, greedy, and structurally untruthful about what it is for; it is not deceptive about what it does. The distinction is the report's central finding and is defended at length in §9 and §10.

(c) "Is it a virus / a stealer?" — Not by capability. Four components, ~2 400 functions, ~90 KB of application code, and not one credential-store reader, cookie reader, wallet reader, keylogger, clipboard reader, or hardware-fingerprinting primitive was found (§9.2, F‑26). Where a component reads something about the user, it reads a username it already knows from the environment and an IP address it already knows from the socket. A stealer would not be so modest.

(d) "Should I run it?" — Only for the executor, and only on a machine where you accept the four consequences in §12.1. The single most useful action available to the recipient is to read the offer page and decline the offers, not to avoid the wrapper. The second most useful is to treat the executor itself as the untrusted component, because it is the one that runs as Administrator, injects into Roblox, and fetches code from the internet on a schedule the wrapper does not control.

13.3 Final judgement

The Subject is not the interesting artefact. It is a well-made, honest, uninteresting decision point — the place where a free piece of software meets a commercial model that has been waiting for it. Judged as software, it is better than most: no packer, a real signature, a per-monitor-aware UI, a typed configuration model, a documented (if inconsistent) consent record, and a clean separation between "the product" and "the offers". Judged as media, it is a lesson: it can be described in 775 KB, signed in 8 072, and monetised in 16.7 MB, and the recipient of all three will experience only one click, one wizard and one progress bar.

The exploit is the interesting artefact, and the report's most durable finding is the distance between them:

the wrapper is the only component of the chain that was engineered for the user, and it is the only component that does not contain a single line of the thing the user wanted.

What the user wanted is the Xeno engine: 2.36 MB of Luau, a bytecode-patch watcher with restore thresholds because Roblox overwrites its own bytecode on every teleport, a remote 248-byte configuration that moves the patch sites with Roblox's version, and a Discord presence whose join secret reaches back through the presence card into a physical server. That engine is genuinely interesting, genuinely maintained, and free. Everything between the user and it — the wrapper, the wizard, the EULA, the offers, the funnel, the bandwidth — is the price of it being free, expressed as an externality the user did not price.

13.4 What would change the assessment

Three falsifiers, in order:

  • OQ‑4 resolves to opt_in-without-consent. If Glory calls hgsdk_opt_in on install and never calls hgsdk_request_consent, then the offer's "It can be disabled at any time in the settings" is still true and the consent model collapses to opt-out. That would move F‑17/F‑18 from Moderate‑High to High and would make the "free service funds itself" framing, strictly, a default-on donation. Current assessment: genuinely unresolved; the strings permit both (OQ‑4). This is the single cheapest experiment in the report and the one most likely to change its headline.
  • A sibling build with a different product name. If a ByteBreaker V2_######## appears with a different eight-digit id and the same nine-string skin, then "Evon" is a publisher, not a one-off repackager — and the re-brand is a product line, not an incident. §6.2 predicts the id will be far from 85892839; the first sighting tests the prediction.
  • The ev (Xeno) CDN going quiet. xeno.now already returns 451 and x3no.pages.dev is flagged Suspected Malware. The engine's remote blob is the one dependency it cannot degrade around: no blob, no attach. A sibling that continues to work after the CDN fails is a sibling with its own engine, not a reskin.

13.5 Closing note (method)

Two techniques account for most of this report and neither is novel.

  1. Read the literals; they are the source. A release MSVC binary with /GS and no obfuscator is a readable document. strings plus RVA attribution plus a disassembler at each hit reconstructs endpoints, field names, error text, identity, business model and culture — faster and more faithfully than decompilation, because literals survive inlining. The Subject ships under 5 KB of them, all plaintext, in two encodings that mark their own semantics.
  2. Ask the server. Two of the three most decisive facts here are not in the file: the descriptor (which names the product) and the offer feed (which names the model). Both are one curl POST away. A static analysis that stops at the binary stops precisely where the design begins — a lesson the subject repeatedly enforced, because every layer of it is split between a shim and a server.

The artefacts that made this cheap were the ones nobody meant to ship: a PDB path from a desktop, a couple of copied comments in a description field, a 2019 Photoshop timestamp in a 32 KB banner, and a joke variable named after a US president.

Provenance is written in the places where the author is not trying to be read.


Appendix A — Complete import table

Ten DLLs, 187 functions. ADVAPI32 is included in full because its four-symbol size is a finding in itself (§3.10).

COMCTL32.dll  (3)
  InitCommonControlsEx  ord#410  ord#413

GDI32.dll  (13)
  CreateFontW  CreatePen  CreateSolidBrush  DeleteObject  GetStockObject
  GetTextExtentPoint32W  LineTo  MoveToEx  RoundRect  SelectObject
  SetBkColor  SetBkMode  SetTextColor

SHELL32.dll  (3)
  SHGetKnownFolderPath  ShellExecuteExW  ShellExecuteW

ole32.dll  (1)
  CoTaskMemFree

urlmon.dll  (2)
  URLDownloadToFileW  URLOpenBlockingStreamA

gdiplus.dll  (12)
  GdipAlloc  GdipCloneImage  GdipCreateFromHDC  GdipDeleteGraphics
  GdipDisposeImage  GdipDrawImageRectI  GdipFree  GdipGetImageHeight
  GdipGetImageWidth  GdipLoadImageFromStream  GdiplusShutdown  GdiplusStartup

WININET.dll  (10)
  HttpOpenRequestA  HttpSendRequestA  InternetCloseHandle  InternetConnectA
  InternetCrackUrlA  InternetOpenA  InternetOpenUrlA  InternetOpenUrlW
  InternetOpenW  InternetReadFile

ADVAPI32.dll  (4)                      <-- registry only; still implies no DPAPI, no crypt, no tokens
  RegCloseKey  RegCreateKeyExW  RegGetValueW  RegSetValueExW

USER32.dll  (36)
  BeginPaint  CreateWindowExW  DefWindowProcW  DestroyWindow  DispatchMessageW
  DrawTextW  EnableWindow  EndPaint  FillRect  FrameRect  GetClientRect  GetDC
  GetMessageW  GetSystemMetrics  GetWindow  GetWindowLongPtrW  GetWindowLongW
  GetWindowTextLengthW  GetWindowTextW  InvalidateRect  LoadCursorW  MessageBoxW
  PostMessageW  PostQuitMessage  RegisterClassExW  RegisterClassW  ReleaseDC
  SendMessageW  SetClassLongPtrW  SetWindowLongW  SetWindowPos  SetWindowTextW
  ShowWindow  TrackMouseEvent  TranslateMessage  UpdateWindow

KERNEL32.dll  (103)   -- application-visible subset
  CreateProcessW  CreateThread  WaitForSingleObject  GetExitCodeProcess
  GetTempPathW  GetTempFileNameW  CreateDirectoryW  DeleteFileW
  GetFileAttributesW  FindFirstFileExW  FindNextFileW  FindClose
  GetModuleFileNameW  GetModuleHandleW  GetCommandLineW  GetEnvironmentStringsW
  SetEnvironmentVariableW  LoadLibraryExW  GetProcAddress  FreeLibrary
  CreateFileW  ReadFile  WriteFile  GetFileSizeEx  SetFilePointerEx
  SetEndOfFile  GetFileType  GetFileAttributesExW  SetFileAttributes…(n/a)
  Sleep  GetTickCount…(n/a)  QueryPerformanceCounter  QueryPerformanceFrequency
  GetSystemTimeAsFileTime  GetCurrentProcessId  GetCurrentThreadId  GetLastError
  SetLastError  IsDebuggerPresent  IsProcessorFeaturePresent  SetUnhandledExceptionFilter
  UnhandledExceptionFilter  TerminateProcess  ExitProcess  ExitThread  GetLastError
  MultiByteToWideChar  WideCharToMultiByte  GetACP  GetOEMCP  GetCPInfo
  IsValidCodePage  IsValidLocale  GetUserDefaultLCID  GetLocaleInfoW  GetLocaleInfoEx
  GetDateFormatW  GetTimeFormatW  GetTimeZoneInformation  EnumSystemLocalesW
  CompareStringW  CompareStringEx  LCMapStringW  LCMapStringEx  GetStringTypeW
  GetDateFormatEx…(n/a)  GetStartupInfoW  GetStdHandle  GetConsoleMode
  GetConsoleOutputCP  ReadConsoleW  WriteConsoleW  FlushFileBuffers  SetStdHandle
  GetProcessHeap  HeapAlloc  HeapFree  HeapReAlloc  HeapSize
  AcquireSRWLockExclusive  ReleaseSRWLockExclusive  InitializeCriticalSectionEx
  EnterCriticalSection  LeaveCriticalSection  DeleteCriticalSection
  InitializeSListHead  WakeAllConditionVariable  FlsAlloc  FlsFree  FlsGetValue
  FlsSetValue  EncodePointer  DecodePointer  SetUnhandledExceptionFilter
  UnhandledExceptionFilter  RaiseException  RtlUnwind  RtlUnwindEx  RtlCaptureContext
  RtlLookupFunctionEntry  RtlPcToFileHeader  RtlVirtualUnwind  VirtualProtect
  GetModuleHandleExW  FreeLibraryAndExitThread  GetEnvironmentStringsW
  FreeEnvironmentStringsW  SetEnvironmentVariableW  CreateThread  GetCurrentProcess
  TerminateProcess  GetLastError  GetProcAddress

  (103 total; the 15 marked "…(n/a)" or omitted above are UCRT/CRT support symbols that are
   imported but have no bearing on behaviour — counted, not itemised.)

Appendix B — Literal pool (selector)

Complete strings are in the companion file. RVA values are image-relative virtual addresses and are stable across copies of the identical image.

B.1 ASCII — wire / diagnostics (2.2 KB total, 0xA01A8 – 0xA0BC0)

0xA01D8  Unknown exception
0xA01F0  bad array new length
0xA0208  bad cast
0xA0218  string too long
0xA0228  map/set too long
0xA0240  [json.exception.
0xA0258  , column
0xA0268   at line
0xA0274  :
0xA0278  iostream
0xA0288  bad locale name
0xA0298  false
0xA02A0  true
0xA02A8  ios_base::badbit set
0xA02C0  ios_base::failbit set
0xA02D8  ios_base::eofbit set
0xA02F0  RegCreateKeyEx failed.
0xA0308  Cannot write string value: RegSetValueEx failed.
0xA0350  Cannot get string value: RegGetValue failed.
0xA0408  request=offers&channel=
0xA0420  https://dlsft.com/service.php?request=offers
0xA0450  started
0xA0458  https://dlsft.com/callback.php
0xA0478  &action=
0xA0484  &id=
0xA0490  channel=
0xA0578  request=getDownloadInfo&name=
0xA0598  channel
0xA05A0  id
0xA05A4  userID
0xA05B0  &comma;
0xA05B8  title
0xA05C0  description
0xA05D0  filename
0xA05DC  logo
0xA05E8  payload
0xA06C8  Mozilla/5.0
0xA06D8  |n
0xA06E0  InternetOpenA failed:
0xA06F8  Failed to parse URL:
0xA0710  InternetConnectA failed:
0xA072C  POST
0xA0738  HttpOpenRequestA failed:
0xA0758  Content-Type: application/x-www-form-urlencoded
0xA0788  HttpSendRequestA failed:
0xA07A8  Error reading from server:
0xA08FC  offers
0xA0904  name
0xA0910  privacy
0xA0918  eula
0xA0920  registryChecks
0xA0930  root
0xA0960  path
0xA0968  value
0xA0970  fileChecks
0xA09A0  targetGeos
0xA09AC  link
0xA09B4  params
0xA0B80  completed

B.2 UTF‑16 — interface, host artefacts, identity (2.4 KB total, 0xA0340 – 0xA0C70)

0xA0340  <fail>
0xA0380  https://filedm.com/privacy.php
0xA03C0  https://filedm.com/terms.php
0xA04A8  C:\Program Files
0xA04D0  https://www.dlsft.com/geo/
0xA0518  .exe
0xA0530  https://www.dlsft.com/service.php
0xA05EE  adreject
0xA0608  glory
0xA0618  start
0xA0628  Download Completed
0xA0650  Download Failed
0xA0670  Open Folder
0xA0688   MB
0xA0690   MB /
0xA06A0  Received
0xA06B8  error
0xA07C8  LOCALAPPDATA
0xA07E8  SOFTWARE\
0xA0800  installed
0xA0818  SOFTWARE\AdReject
0xA0840  Consent
0xA0850  SOFTWARE\Glory
0xA0870  ADR
0xA0878  " --silent
0xA0898  GLY
0xA08A0  SCI_multipart
0xA08C0  \file.txt
0xA08D8  open
0xA08E8  &s=
0xA08F0  ?id=
0xA0938  HKEY_LOCAL_MACHINE
0xA0980  %ProgramFiles%
0xA09C0  Installer
0xA09E0  Window Class Registration Failed! Error code:
0xA0A40  Error
0xA0A50  Calibri
0xA0A60  Segoe UI
0xA0A78  STATIC
0xA0A88  BUTTON
0xA0AA0  Next
0xA0AC0   Download Manager.
0xA0AE8  Welcome to the
0xA0B08  Privacy Policy
0xA0B28  EULA
0xA0B38   - Optional Offer
0xA0B60  Decline
0xA0B70  Accept
0xA0B90  Thanks for using our Download Manager
0xA0BE0  Open File
0xA0BF8  Download
0xA0C10  msctls_progress32
0xA23D0  textbox
0xA23E0  App
0xA2428  EDIT
0xA2438  COMBOBOX
0xA2450  Finish
0xA2480  PremierOpinion
0xA24A0  ContentI3.exe
0xA25A0  Survey
0xA25B0  Please complete the short survey.
0xA25F4  po-2
0xA2660  BundleOfferActionUid
0xA2690  SOFTWARE\InstallUnion\1538
0xA26D0   helps us provide this download for free. You can support us by accepting their offer.
0xA2780  SysLink
0xA27E0  SurveyWinClass
0xA2800  SurveyHost

B.3 Engine (Xeno) — selected identity strings

C:\Users\rizve\OneDrive\Desktop\Projects\Xeno\build\Release\net9.0-windows8.0\Xeno.pdb
rizve@xeno.onl
.xeno      (Discord handle, prefixed '.')
https://xeno.now
https://pastebin.com/raw/A7n0vJtR
https://x3no.pages.dev/ep.zst.enc
Xeno Server thread: {:#x}
Xeno Caching thread: {:#x}
WebSockets Cleaner thread: {:#x}
Client Checker thread: {:#x}
A Xeno resource was unable to load, will abort immediately.
Using Xeno
Xeno/
%XENO_VERSION%
%XENO_SERVER_PORT%
[{} ({})][::BytecodePatchWatcherT]: Threshold reached. Workers failed to call restore
   (Automatically restoring to prevent crash)
QueueOnTeleport
Loaded {} Cached queued scripts for QueueOnTeleport
HoneygainSdk: missing export          (Glory.exe)

Appendix C — Cryptographic register

C.1 Sample hashes

Object Algorithm Value
Subject SHA‑256 bd1f1619f850be8c84ae0acdd0bb7ff3d88b8cb5647877c078d7805d7a55582f
Subject SHA‑1 932e7cce8f6cde321f2288ccaaeb8ae5d84ab26d
Subject MD5 280603b04aeadbb0d42f6043354923cd
Subject Authentihash SHA‑256 47cb2bce143f52236604570e76081fa7bbf49c27955cba8be9ea4780f1d1b1c1
Subject Authentihash SHA‑1 4993f6423a4650be8d0190045aff69b8463f34da
Subject imphash b3eb077da70cf3699e3dc8149d29c2be
Subject Rich key 0x9176119f
Product archive SHA‑256 aa3fc2c089a8f039b2198d6f64e4e727e1631c7e18fe084429b1b491330ad85f
Product archive Size 16,787,263 B
Control-plane blob (both mirrors) SHA‑256 e4b57255b6b01b12a7a729d3155294600ef2d43af2ae13d6e1126f15a4126381
Control-plane blob Size 276 B (identical after base64 decode)
Glory package SHA‑256 1b67e6b657b356d8e0658de2cb20010e0520542e8fd9b2392c5b5ee93263101c
Glory package Size 12,579,642 B
Glory.exe SHA‑256 3111a67c5708f2caf49fe362ff48319e094da50bb2cc60a75aab62c162525068
hgsdk.dll SHA‑256 2720d215168108700767f8e9b57819d55ca139d69d8d828186da186850a37c6c
sciter.dll SHA‑256 8ec04a9f38c3cbc376d11330d9022b830dafa0d27cd4ac89f4ab87ed8b883270
AdReject package SHA‑256 59358bb52a3985faae054c78bae0558112894bf5027f4a6dce34e3acba726e0a
AdReject package Size 574,773 B
AdReject.exe SHA‑256 e943c047bc24ad68a0ce23856ce954a8c027c5d7f44c6bff3fefc6a29307af54
WinDivert.dll SHA‑256 c1e060ee19444a259b2162f8af0f3fe8c4428a1c6f694dce20de194ac8d7d9a2
WinDivert64.sys SHA‑256 8da085332782708d8767bcace5327a6ec7283c17cfb85e40b03cd2323a90ddc2
aichat_2.0.3_x64-setup.exe SHA‑256 6430a772eb047467aa0c3e4bc964719dbf85b9adc58e63738a8652099ba6f8ec
Subject SHA‑256 aa3fc2c089a8f039b2198d6f64e4e727e1631c7e18fe084429b1b491330ad85f (verify)
Subject SHA‑1 932e7cce8f6cde321f2288ccaaeb8ae5d84ab26d

C.2 Signature material

SignedData.version              1
digestAlgorithms               sha256
encapContentInfo.eContentType  1.3.6.1.4.1.311.2.1.4        (SPC_INDIRECT_DATA)
signerInfo.sid                 issuerAndSerial
signerInfo.digestAlgorithm     sha256
signerInfo.signatureAlgorithm  1.2.840.10045.2.1             (id-ecPublicKey)
signedAttrs                    content-type, message-digest, 1.3.6.1.4.1.311.2.1.12 (SPC_SP_OPUS),
                               1.3.6.1.4.1.311.2.1.11 (SPC_STATEMENT_TYPE) = 1.3.6.1.4.1.311.2.1.21
unsignedAttrs                  counter-signature (RFC-3161 token)

C.3 Timestamp token

subject  CN=DigiCert SHA256 RSA4096 Timestamp Responder 2026 1
         O=DigiCert, Inc., C=US
         serial 11048481613454637687985428644217624629
validity 2026-08-05 .. 2037-11-04
issuer   CN=DigiCert Trusted G4 TimeStamping RSA4096 SHA256 2025 CA1
         serial 18316726489210163671024321415498844550
root     CN=DigiCert Trusted Root G4
         issuer CN=DigiCert Assured ID Root CA

Appendix D — Methodology and tooling

D.1 Environment

Host      Linux x86-64 (Ubuntu 24.04 base)
Date      2026-11-16 → 2026-11-17
Egress    GB (residential/DC unknown to the server; reported by /geo/ as "GB")
Python    3.12.3   (venv at _work/venv for analysis dependencies)
Java      Temurin 21.0.13 (JRE; required by the decompiler)

D.2 Tooling, by task

Task Tool Version Notes
PE parsing, imports, version resources pefile 2024.8.26 primary structural tool
Signature / ASN.1 asn1crypto 1.5.1 PKCS#7, SignedData, extensions
Signature helpers signify — installed for cross-check; asn1crypto used
Disassembly radare2 5.5.0 entry, xrefs, /r reference search
Decompilation Ghidra 11.2.1 (Ghidra_11.2.1_PUBLIC_20241105) headless, custom DumpDecomp.java post-script
Ghidra runtime Temurin JDK 21.0.13 JAVA_HOME_OVERRIDE + java on PATH (both required)
Managed metadata dnfile current BSJB, TypeDef/MethodDef/AssemblyRef
Crypto experiments cryptography 50.0.1 AESGCM, ChaCha20Poly1305
Pattern matching YARA 4.5.0 rule compilation + validation
Archive handling 7-Zip 16.02 p7zip-full; NSIS and Velopack ZIPs
HTTP curl 8.x all endpoint interrogation
Strings GNU binutils strings 2.42 ASCII and UTF‑16LE modes
Compressed-stream filter binwalk 2.3.4 installed; used only for magic confirmation
Unpacker probe UPX 4.2.2 negative result (§3.2)

D.3 Decompilation configuration

Ghidra was run headless on the default x86‑64 PE loader with no PDB, default analysis, and a per-function decompile timeout of 120 s. A post-script iterated FunctionManager and emitted every non-external, non-thunk function with an /* === NAME @ ADDR size=N === */ banner:

DecompInterface ifc = new DecompInterface();
ifc.setOptions(new DecompileOptions());
ifc.openProgram(currentProgram);
for (Function f : currentProgram.getFunctionManager().getFunctions(true)) {
    if (f.isExternal() || f.isThunk()) continue;
    DecompileResults r = ifc.decompileFunction(f, 120, monitor);
    …
}

Output: 1 654 functions, 122 242 lines, zero decompilation failures. The per-function file split (fn/NAME__ADDR.c) is what made the application layer navigable; 1 654 functions in one 3.5 MB file are not.

D.4 Statistical caveats

  • The decompiled C is pseudo-C. Types are inferred; structure layout is approximate; this-relative member access is rendered as pointer arithmetic; inlined std::wstring accessors appear as raw byte offsets. Where this report quotes "g_id.size()" or "& 0xFF", the semantics were confirmed by cross-checking the format strings and the runtime behaviour of the algorithm (§5.4), not by trusting Ghidra's types.
  • Function counts are not code-size metrics: 1 654 functions include every CRT thunk, every exception funclets and every COMDAT-folded template instantiation. The ~35 application functions figure is a hand count over 0x140003000–0x140035000, not a filter.
  • Entropy figures are Shannon entropy over the section's bytes; they distinguish stored from compressed but say nothing about encrypted. §6.5's arguments about encrypted strings rest on the absence of higher-entropy islands, not on the section mean.

Appendix E — Corrections log and confidence ledger

This assessment had a Part I (the wrapper, §1–§5) and a Part II (everything behind it, §6–§13). Part II corrected Part I in eight places. The corrections are recorded because a reader who saw only Part I would otherwise hold eight beliefs that are now known to be false, and because the pattern of the errors is itself instructive: every one is a premature closure — a question answered before the evidence that would have answered it was collected.

# Part I said Part II established Class of error Lesson
1 SCI_multipart was "unresolved — may be dead code from a shared header" It is the hotlink salt (§5.4); without it no product can be fetched dead code assumed "unused string" is an oxymoron in a linker; strings are used or absent
2 ?id= and &s= were "two unused 3-byte literals" They build every product URL (§5.4) as above same
3 The _85892839 suffix was "medium‑high confidence inference" It is a server-side row key, echoed as "id" (§6.2) inference under‑stated as guess a live endpoint costs one POST and converts a guess into a fact
4 The encrypted blob was "258 B = 12 + 230 + 16" 276 B = 12 + 248 + 16, and both mirrors are byte-identical (§6.13) arithmetic slip on base64 length base64 length is 4·ceil(n/3), not 4n/3; recompute, then reconcile
5 The product's vendor was "unknown"; "AN Soft" was the wrapper's vendor AN Soft is the wrapper, the CDN, the platform and 2 of the 5 offers; the product is rizve's Xeno, re-skinned by Evon (§6.5, §8.1) single-entity attribution one signature does not imply one author; draw the whole graph
6 ContentI3.exe / PremierOpinion "appear in the wrapper" (string level) They are one row of the live feed, with 28-country targeting, -c:1538 -t:InstallUnion arguments, and a verified comScore consent text (§7.6) string ↔ data conflation a string is a capability; the feed is the fact
7 "The offers are what the wrapper shows" The offers are what the server sold that day; the wrapper is unit-agnostic and targetGeos-gated (§7.10) static assumption the feed is the product; the binary is the shopfront
8 The Subject "installs five offers" It installs at most one or two, and only on Accept; the rest are impressions it records (§10.4 S‑1) language distinguish impression, consent and installation — three different events, three different parties paid

Net effect of the corrections: none of them changed a verdict; four changed severity (up: 5, 6; down: 1, 8); two changed attribution (5, 8); one was purely arithmetic (4). That ratio — verdicts stable, severities moving — is what one expects from an assessment whose conclusions are structural and whose numbers are provisional, and it is offered as a calibration datum for the confidence ratings in §2.3.


Appendix F — Open question register

ID Question Why unresolved Method to resolve Effort Impact if resolved
OQ‑1 The eight-digit identifier is a product key, not a publisher key — so which other ids in 8589xxxx belong to this publisher? 701-id scan (§6.2) found a single product in that window; the allocation strategy is unknown scan ±10⁵ ids at 8–12 threads; cluster by payload host and by channel 30 min maps the publisher's catalogue, and therefore its economics
OQ‑2 Is the directfiledl host shared across publishers (i.e. a platform CDN) or dedicated? the host serves one getfile script; other publishers' links not observed fetch descriptors for products from other FileDM uploaders seen in the wild; diff payload hosts 2 h distinguishes platform infrastructure from publisher infrastructure
OQ‑3 What is the 32-byte AES key for the remote configuration (§6.14)? 9.37 M static candidates eliminated; key is derived break on BCryptDecrypt; dump key + IV; decrypt both mirrors 2–4 h opens the engine's instruction set; may contain API list, features, kill-switches
OQ‑4 Does Glory call hgsdk_request_consent at all, or only hgsdk_opt_in? both exports are resolved; call sites not statically ordered xref the two GetProcAddress results; read the branch; or first-run UI capture 3–5 h converts F‑18 from "consent API exists" to "consent UX is implemented or not"; re-rates F‑17
OQ‑5 Is the installation id persisted anywhere client-side? 16-hex value is held in memory; no file/registry literal observed for it run twice; compare /geo/ responses; watch with procmon 1 h determines whether the id is session or device scope → §9.4
OQ‑6 Does the wrapper ever re-request getDownloadInfo after first bootstrap (i.e. is the product pinned)? request count observed only in a single session run for 7 d; count requests per launch 7 d determines whether a product swap (§10.2 S‑8) is possible without re-download
OQ‑7 What exactly does ContentI3.exe do at install and at rest? not retrieved (§2.1) download 4xx–8xx KB; unpack; install in VM under -c:1538 -t:InstallUnion; Procmon + sc query + WMI 4 h the only offer whose purpose is observation; currently characterised from its description alone
OQ‑8 Is the "best mag" capture EULA retained in the deliverable or fetched? Editor/index.html references the string; no local file check WebView2 cache at %LOCALAPPDATA%\…\EBWebView after first run 30 min distinguishes "consent to local file" from "consent to mutable remote text" — the strongest single finding if remote
OQ‑9 Does the Quorum.dll engine detect its own modification (anti-tamper)? not tested; no hash-check literal observed patch the export table / rename Initialize; run; observe 2 h determines whether the engine is replaceable by a third party (F‑09's corollary)
OQ‑10 Are the !DE exclusions dynamic (server-configurable) or static (build)? two products agree; no feed for an excluded country was obtained sample /service.php?request=offers from DE and non-DE egress; diff; or request channel variants 1 h decides whether Germany is a policy or a build option (§7.10)
OQ‑11 Does the wrapper phone home on every launch or only the first? callback.php?act=started observed; run count not compared 10 launches, count beacons 15 min the funnel is "per installation" or "per session"; different denominators
OQ‑12 Is directfiledl.com shared with the other 4 offer hosts or dedicated? only glory.tv/adreject.com manifests were read fetch all five manifests; compare package_url hosts 10 min infrastructure seam (F‑15)

Appendix G — Acquisition notes

Every artefact in this report was obtained from a public, unauthenticated endpoint with an ordinary HTTP GET/POST and a Mozilla/5.0 user agent. Nothing was purchased, registered, gated, rate-limited (until the 701-id scan) or supplied by a third party. Reproduction commands are in §35 of the companion file; this appendix records only the provenance of each artefact and the local size, so that a reader can confirm the same object was in hand.

Artefact Acquired from Method Size (B) SHA‑256 (prefix)
Subject ByteBreaker V1_85892839.exe provided — 775,048 bd1f1619…
Offer bundle POST dlsft.com/service.php?request=offers form 8 887 —
Product descriptor POST …&request=getDownloadInfo&name=ByteBreaker V1_85892839 form 232 —
bytebreakerV1.zip GET http://directfiledl.com/getfile?id=85892839&s=6f75a7cd GET 16,787,263 aa3fc2c0…
ep.zst.enc (mirror A) GET https://x3no.pages.dev/ep.zst.enc GET 276 e4b57255…
A7n0vJtR (mirror B) GET https://pastebin.com/raw/A7n0vJtR → b64 GET 344 → 276 e4b57255…
PO.jpg GET https://dlsft.com/PO.jpg GET 32,755 —
privacy.php / terms.php GET https://filedm.com/… GET 9 235 / 7 635 —
Glory-1.0.7-windows-x64.zip GET https://glory.tv/releases/windows-x64/Glory-1.0.7-windows-x64.zip GET 12,579,642 1b67e6b6…
AdReject-1.0.9-windows-x64.zip GET https://adreject.com/releases/windows-x64/AdReject-1.0.9-windows-x64.zip GET 574,773 59358bb5…
aichat_2.0.3_x64-setup.exe GET https://github.com/aichatdl/…/aichat_2.0.3_x64-setup.exe GET 3,807,352 6430a772…
latest.manifest ×2 GET https://{glory.tv,adreject.com}/releases/windows-x64/latest.manifest GET 1 077 / 1 300 —
latest.json (Tauri) GET https://github.com/aichatdl/…/releases/latest/download/latest.json GET 1 200 —

All retrieval was read-only and idempotent except:

  • 701 getDownloadInfo probes (a global hit count, not per-install — so they cannot be attributed to a host in any publisher-facing statistic);
  • 49 requests to directfiledl.com/getfile with candidate s values, of which 1 succeeded and returned the archive;
  • 6,000 candidate s values computed locally, not requested (the strap-line sweep is arithmetic, not HTTP);
  • 27 DNS resolutions.

The first three are worth stating in a report because they define the observation footprint the analysis leaves on the vendor's side. That footprint is:

  dlsft.com       : ~760 requests  in ~30 h   (1 IP, 1 UA, 1 ASN)
  directfiledl.com:   49 requests  in  ~1 h
  filedm.com      :    2 requests
  pastebin.com    :    1 request
  x3no.pages.dev  :    2 requests

For a service that measures installs, 760 requests from one ASN in thirty hours is noise. No analysis in this report was distorted by self-observation, and the vendor did not react to the probing during the window.

Analyst's disclosure. No conflict of interest. The Subject was not executed on the analyst's host. All remote endpoints were fetched over a residential United Kingdom connection; no VPN, proxy or geo-spoofing was used for any measurement reported in this document, and where the egress location matters to a conclusion (country filtering, §5.3) the value GB is stated with the result.


End of report. Companion working notes, decompilation (decomp.c, fn/), literal dumps, rules and machine-readable indicators are in ByteBreaker_V1_85892839_ANALYSIS.md and _work/.

Import Address Table (10 DLLs)
  • COMCTL32.dll
    ord_410, ord_413, InitCommonControlsEx
  • KERNEL32.dll
    GetModuleHandleW, SetEndOfFile, WriteConsoleW, HeapSize, CreateFileW, SetStdHandle, GetProcessHeap, SetEnvironmentVariableW, FreeEnvironmentStringsW, GetEnvironmentStringsW, GetCommandLineW, GetCommandLineA, FindNextFileW, FindFirstFileExW, FindClose, GetFileAttributesW, HeapReAlloc, GetACP, IsValidCodePage, ReadConsoleW, ReadFile, GetFileAttributesExW, EnumSystemLocalesW, GetUserDefaultLCID (+79 more)
  • USER32.dll
    MessageBoxW, SetWindowPos, DestroyWindow, PostMessageW, DefWindowProcW, CreateWindowExW, GetWindowTextW, EnableWindow, EndPaint, BeginPaint, InvalidateRect, SendMessageW, PostQuitMessage, DrawTextW, GetClientRect, SetWindowLongW, LoadCursorW, TranslateMessage, FrameRect, TrackMouseEvent, DispatchMessageW, ShowWindow, RegisterClassExW, GetWindowLongPtrW (+12 more)
  • GDI32.dll
    GetTextExtentPoint32W, SelectObject, CreateFontW, GetStockObject, SetTextColor, SetBkMode, LineTo, CreatePen, MoveToEx, SetBkColor, DeleteObject, CreateSolidBrush, RoundRect
  • ADVAPI32.dll
    RegGetValueW, RegSetValueExW, RegCloseKey, RegCreateKeyExW
  • SHELL32.dll
    SHGetKnownFolderPath, ShellExecuteExW, ShellExecuteW
  • ole32.dll
    CoTaskMemFree
  • urlmon.dll
    URLDownloadToFileW, URLOpenBlockingStreamA
  • gdiplus.dll
    GdiplusStartup, GdiplusShutdown, GdipGetImageHeight, GdipCloneImage, GdipAlloc, GdipLoadImageFromStream, GdipGetImageWidth, GdipDeleteGraphics, GdipCreateFromHDC, GdipFree, GdipDisposeImage, GdipDrawImageRectI
  • WININET.dll
    HttpOpenRequestA, InternetCrackUrlA, InternetOpenUrlA, InternetOpenA, InternetCloseHandle, HttpSendRequestA, InternetReadFile, InternetOpenUrlW, InternetOpenW, InternetConnectA
Extracted Indicators of Compromise (92)
TypeIndicatorContext
sha256 bd1f1619f850be8c84ae0acdd0bb7ff3d88b8cb5647877c078d7805d7a55582f Subject wrapper (sample)
sha1 932e7cce8f6cde321f2288ccaaeb8ae5d84ab26d Subject wrapper
md5 280603b04aeadbb0d42f6043354923cd Subject wrapper
authentihash-sha256 47cb2bce143f52236604570e76081fa7bbf49c27955cba8be9ea4780f1d1b1c1 signature-independent hash
authentihash-sha1 4993f6423a4650be8d0190045aff69b8463f34da signature-independent hash
imphash b3eb077da70cf3699e3dc8149d29c2be family pivot
sha256 aa3fc2c089a8f039b2198d6f64e4e727e1631c7e18fe084429b1b491330ad85f delivered product archive bytebreakerV1.zip (16,787,263 B)
sha256 e4b57255b6b01b12a7a729d3155294600ef2d43af2ae13d6e1126f15a4126381 276 B engine remote-config blob (identical on both mirrors)
sha256 1b67e6b657b356d8e0658de2cb20010e0520542e8fd9b2392c5b5ee93263101c Glory-1.0.7-windows-x64.zip (12,579,642 B)
sha256 3111a67c5708f2caf49fe362ff48319e094da50bb2cc60a75aab62c162525068 Glory.exe 1.0.7 (AN Soft)
sha256 2720d215168108700767f8e9b57819d55ca139d69d8d828186da186850a37c6c hgsdk.dll 1.5.0 (Honeygain SDK)
sha256 8ec04a9f38c3cbc376d11330d9022b830dafa0d27cd4ac89f4ab87ed8b883270 sciter.dll (Glory UI)
sha256 59358bb52a3985faae054c78bae0558112894bf5027f4a6dce34e3acba726e0a AdReject-1.0.9-windows-x64.zip (574,773 B)
sha256 e943c047bc24ad68a0ce23856ce954a8c027c5d7f44c6bff3fefc6a29307af54 AdReject.exe 1.0.9 (AN Soft)
sha256 c1e060ee19444a259b2162f8af0f3fe8c4428a1c6f694dce20de194ac8d7d9a2 WinDivert.dll (as shipped)
sha256 8da085332782708d8767bcace5327a6ec7283c17cfb85e40b03cd2323a90ddc2 WinDivert64.sys (as shipped)
sha256 6430a772eb047467aa0c3e4bc964719dbf85b9adc58e63738a8652099ba6f8ec aichat_2.0.3_x64-setup.exe
filename ByteBreaker V1_85892839.exe convention: <product> V<major>_<8-digit server id>.exe
filename bytebreakerV1.zip product archive (nullsoft 7-zip wrapper of Xeno)
filename Quorum.dll / QuorumAPI.dll engine / managed bridge inside archive
filename decompiler.exe Medal (Rust) Luau decompiler inside archive
filename version.json / current_version.txt / lastKey.txt 1.4.816 / 1.2.0 / "keysystemremoved"
path %TEMP%\ADR<hex4>.exe brand-A staged installer (GetTempFileNameW, deleted after run)
path %TEMP%\GLY<hex4>.exe brand-B staged installer
path %LOCALAPPDATA%\adreject\file.txt brand-A state
path %LOCALAPPDATA%\glory\file.txt brand-B state
path %ProgramFiles%\Glory\ Glory.exe + hgsdk.dll + sciter.dll
path %ProgramFiles%\AdReject\ AdReject.exe + WinDivert.dll + WinDivert64.sys + blocklist.txt + lists\
path %ProgramFiles%\PremierOpinion\ comScore panel (pmropn.exe)
path %ProgramData%\Glory\ capture library (SQLite), thumbnails, logs
path %APPDATA%\aichat\ Tauri AI client
path %LOCALAPPDATA%\Google\Chrome\User Data\ NOT accessed by this family
registry HKCU\SOFTWARE\InstallUnion\1538\BundleOfferActionUid REG_SZ offer URL/uid; best single family pivot
registry HKCU\SOFTWARE\AdReject\Consent REG_DWORD 1, written before child launch
registry HKCU\SOFTWARE\Glory\Consent REG_DWORD 1, written before child launch
registry HKCU\SOFTWARE\Classes\OperaGXStable / OperaStable (HKCU+HKLM, value "URL Protocol") suppression checks, read-only
registry HKCU\SOFTWARE\OperaGX\installed ; HKCU\SOFTWARE\AIChat\installed suppression checks, read-only
process "%TEMP%\ADR<hex4>.exe" --silent CREATE_NO_WINDOW, cwd %TEMP%, parent blocks until exit
process "%TEMP%\GLY<hex4>.exe" --silent -c:1538 -t:InstallUnion server-supplied params appended unquoted
process "<product>V<n>_<8digits>.exe" parent wrapper name, 8-digit numeric suffix
window class "Installer" / title "Installer" 692x492 WS_POPUP no icon, generic default application icon
named-object Global\AdReject.Update.Request.v1 AdReject CLI<->GUI control channel
named-object WinDivert 2.x / WinDivert64.sys service kernel packet-divert driver
url https://dlsft.com/service.php?request=offers POST, body request=offers&channel=<c>
url https://dlsft.com/service.php?request=getDownloadInfo&name=<name> POST, resolves by trailing 8-digit id
url https://dlsft.com/callback.php?channel=<c>&act=started completed&id=<i> | vendor funnel receipt
url https://www.dlsft.com/geo/ returns "<16 hex>,<ISO-3166>"
url http://post.securestudies.com/TapAction.aspx?&uid=<id>campaign_id=1538&tpi=InstallUnion&action_id=<n> cleartext HTTP; action_id 0,1,2,3,4,8,9,17
url https://dpd.securestudies.com/dpdv3.aspx?campaignid=1538&co=<v <Blocked>> | post-download pixel
url https://dlsft.com/PO.jpg 413x30 "promo offer" strip, EXIF 2019-10-04 Photoshop CC 2019 (Mac)
url https://filedm.com/privacy.php ; https://filedm.com/terms.php linked legal pages (FileDM/AN Soft brand)
url http://directfiledl.com/getfile?id=85892839&s=6f75a7cd product archive; s = crc32(id XOR "SCI_multipart")
url https://glory.tv/releases/windows-x64/latest.manifest Velopack/Squirrel-style update index
url https://adreject.com/releases/windows-x64/latest.manifest same
url https://github.com/aichatdl/aichat-app-tauri/releases/download/v2.0.3/aichat_2.0.3_x64-setup.exe offer payload
url https://post.securestudies.com/packages/PI1032/ContentI3.exe PremierOpinion installer
url https://raw.githubusercontent.com/ShadowWhisperer/BlockLists/master/Lists/Ads AdReject managed blocklist
url https://pastebin.com/raw/A7n0vJtR Xeno engine remote-config mirror A (base64)
url https://x3no.pages.dev/ep.zst.enc mirror B (276 B, byte-identical)
url https://xeno.now product site; currently HTTP 451
url http://my.aichat.rocks:9005 AIChat cleartext backend, non-standard port
url https://www.ovardu.com/6R9H6BD/KMZXBTT/?sub1=FileDM&sub=default OperaGX CPA redirect
domain dlsft.com, www.dlsft.com Cloudflare; control plane
domain filedm.com Cloudflare; brand/legal/analytics
domain directfiledl.com Hetzner; data plane, plain HTTP only
domain glory.tv, adreject.com Cloudflare; first-party offers
domain securestudies.com ad/measurement network (post., dpd.)
domain x3no.pages.dev Cloudflare Pages; engine config mirror
domain connect.sdk.honeygain.com honeygain UAB SDK endpoint (proxyware)
domain ovardu.com affiliate tracker
domain aichat.rocks Tauri app backend
ip 116.202.13.47 directfiledl.com A record (Hetzner, AS24940)
ip 2a01:4f8:c012:42ca::1 directfiledl.com AAAA record (Hetzner IPv6)
cert CN=AN Soft, O=AN Soft, L=Bergen op Zoom, ST=North Brabant, C=NL; serial 158467513303841965040394699300328354453; SSL.com Code Signing Intermediate CA ECC R2; 2026-06-11 -> 2027-06-11 shared by wrapper, Glory.exe, AdReject.exe
cert CN=Honeygain UAB, O=Honeygain UAB, L=Vilnius, C=LT, SERIALNUMBER=306103177; DigiCert Trusted G4 Code Signing RSA4096 SHA384 2021 CA1 hgsdk.dll
tsa CN=DigiCert SHA256 RSA4096 Timestamp Responder 2026 1 2037 validity, keeps wrapper verifiable long-term
pdb C:\Users\Samim\Desktop\Installer\x64\Release\Installer.pdb wrapper build provenance
pdb C:\Users\rizve\OneDrive\Desktop\Projects\Xeno\build\Release\net9.0-windows8.0\Xeno.pdb engine identity -> Xeno
pdb C:\Users\Samim\Desktop\Glory\x64\Release\Glory.pdb first-party offer, same developer
pdb C:\Users\Samim\Desktop\Glory\Recorder.cpp ; ScreenshotCapture.cpp source-file leakage
guid 40261B86-E57C-4ACB-BC2F-99184C3D8202 (age 3) RSDS, wrapper
richkey 0x9176119f linker fingerprint
string SCI_multipart hotlink salt (also a family anchor)
string campaign_id=1538&tpi=InstallUnion ad-network campaign
string SOFTWARE\InstallUnion\1538 campaign registry root
string BundleOfferActionUid / ContentI3.exe / PremierOpinion offer vocabulary
string "Security code is not matching: carefully copy and do not change the link." file-host validator response
string HoneygainSdk: missing export Glory/SDK integration error string
string "Your device may function as a network exit node" proxyware consent copy
ssdeep 6144:... withheld — recompute on corpus before publishing
file-size 775048 wrapper
file-size 16787263 product archive
YARA Signature
import "pe"

rule ANSoft_FileDM_InstallUnion_Downloader
{
    meta:
        description   = "AN Soft / FileDM white-label PPI download wrapper (InstallUnion campaign 1538)"
        family          = "AN Soft Downloader"
        author          = "Elera"
        date            = "2026-11-17"
        reference_hash  = "bd1f1619f850be8c84ae0acdd0bb7ff3d88b8cb5647877c078d7805d7a55582f"
        reference_size  = 775048
        imphash         = "b3eb077da70cf3699e3dc8149d29c2be"
        tlp             = "CLEAR"
        scope           = "file"
        note            = "family rule: survives re-brand, re-sign and new 8-digit id"

    strings:
        /* control plane */
        $u_offers   = "https://dlsft.com/service.php?request=offers"            ascii
        $u_callback = "https://dlsft.com/callback.php"                          ascii
        $u_po       = "https://dlsft.com/PO.jpg"                                ascii
        $u_geo      = "https://www.dlsft.com/geo/"                              ascii
        $u_service  = "https://www.dlsft.com/service.php"                      ascii

        /* request shapes */
        $q_offers   = "request=offers&channel="                                 ascii
        $q_dlinfo   = "request=getDownloadInfo&name="                          ascii

        /* telemetry */
        $t_tap      = "post.securestudies.com/TapAction.aspx?"                   ascii
        $t_dpd      = "dpd.securestudies.com/dpdv3.aspx?campaignid=1538&co="     ascii
        $t_camp     = "campaign_id=1538&tpi=InstallUnion"                       ascii wide

        /* persistent state */
        $r_campaign = "SOFTWARE\\InstallUnion\\1538"                            wide
        $r_receipt  = "BundleOfferActionUid"                                    wide
        $r_adr      = "SOFTWARE\\AdReject"                                      wide
        $r_gly      = "SOFTWARE\\Glory"                                         wide

        /* distinctive literals */
        $s_salt     = "SCI_multipart"                                           wide
        $s_po       = "PremierOpinion"                                          wide
        $s_content  = "ContentI3.exe"                                           wide
        $s_po_slot  = "po-2"                                                     wide

        /* interface copy */
        $w_class    = "Installer"                                               wide fullword
        $w_regfail  = "Window Class Registration Failed! Error code: "          wide
        $w_dlmgr    = " Download Manager."                                      wide
        $w_optional = " - Optional Offer"                                       wide
        $w_thanks   = "Thanks for using our Download Manager"                   wide

    condition:
        uint16(0) == 0x5A4D and
        pe.is_pe and
        filesize < 5MB and
        (
            pe.imphash() == "b3eb077da70cf3699e3dc8149d29c2be"
            or
            /* offer pipeline + at least one endpoint */
            ( $q_offers and $q_dlinfo and 1 of ($u_*) )
            or
            /* campaign state: registry root + receipt (+ a brand) */
            ( $r_campaign and $r_receipt and 1 of ($r_adr, $r_gly) )
            or
            /* salt is high-precision: never observed outside this family */
            ( $s_salt and 1 of ($w_*) )
            or
            /* two independent telemetry strings */
            ( $t_camp and ( $t_tap or $t_dpd ) )
            or
            /* stripped/de-duplicated rebuild: count the unique literals */
            ( 3 of ($u_*, $q_*, $r_*, $s_*, $w_*) )
        )
}