OffSeq field investigation 002: J@IL-GPT
Research date: 6 September 2026

This bundle accompanies the investigation of the DarkForums post 423476 and
the GitHub repository kikus662zqdz/j-ilgpt. It contains derived evidence and
defensive analysis tools. It contains no malware executables or archives,
victim credentials, complete forum capture, or runnable implant code.

FILES

evidence.json
  Preserved artifact metadata, distribution provenance, function addresses,
  reconstructed behaviors, network pivots, limitations and YARA validation.
  The eight malware-artifact SHA-256 values identify private preserved files;
  those files are not included here.

artifacts.csv
  Eight artifacts with sizes, SHA-256 values, parent relationships and method.

repository-timeline.csv
  Nine commits and release publication. All timestamps in this CSV are UTC.
  The forum's unlabeled local timestamp is kept separately in evidence.json.

offseq-jailgpt.yar
  Two sample-lineage rules: outer PE log container and recovered inner PE.
  YARA 4.5.4 compilation and 15 focused checks passed: four intended positive
  cases, three benign controls, four truncations and four footer mutations.
  The outer rule requires a PE, including an extracted MSI Binary.p stream.
  The inner rule requires decoded bytes. Neither scans through an unopened
  MSI or decodes the wrapper. The embedded browser DLLs are identified by
  hash; the inner-PE rule is not a DLL rule. No broad false-positive benchmark
  or complete family coverage is claimed.

extract-jailgpt.py
  Offline, exact-hash extraction for the two reviewed outer PE files. Reads
  bytes, reconstructs the inner executable and embedded DLL, verifies their
  expected hashes, and writes .sample analysis files. No network calls,
  child processes, native-code loading or malware execution. Run --help for
  usage. It does not extract the Binary table from the MSI itself. Recovered
  outputs remain malware and must be handled as such; renaming is not a
  deactivation mechanism.
  Six focused checks passed with Python 3.9.6 in isolated stdlib mode: both
  exact reconstructions, two overwrite refusals and two unknown-input refusals.

artifact-chain-en.svg / artifact-chain-lv.svg
artifact-chain-en.png / artifact-chain-lv.png
  English and Latvian figures showing the preserved package relationships.
  Figure source: scripts/research/plot-jailgpt-chain.py in the website repo.

domain-warning-20260906.png
  Unedited screenshot supplied by the user during the investigation. It
  shows a warning when navigating to the fallback domain over HTTP. It is
  not an independent research-team network capture, proof of an accepted
  HTTPS upload, provider attribution, or a confirmed infection.

SHA256SUMS.txt
  Integrity hashes for the public bundle files, excluding this checksum list.

INTERPRETATION

This was static analysis and offline reconstruction. No suspect executable
was run and no extracted command-and-control destination was contacted by
the researchers. Collection, injection, upload and conditional persistence
are behaviors established in code. Their success on a particular victim
was not observed. A download count is not a victim count. A hash match on
disk does not by itself prove execution. Shared infrastructure and copied
README text do not establish a real-world operator or definitive family.

Network URLs are defanged where presented as indicators. The two protocol
markers occur inside encrypted HTTPS bodies and are not plaintext network
signatures. The screenshot's successful challenge widget is not a verdict
that the destination is safe.

Original analysis and derived figures: OffSeq Cybersecurity.