os-eval-backport

Pass

Audited by Gen Agent Trust Hub on Apr 3, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: Indirect Prompt Injection Surface. The skill is designed to ingest and act upon data (logs and retrospectives) from a 'lab repo' that may have been generated by other agents or processes. This external data is considered untrusted.
  • Ingestion points: The skill reads files from <lab-repo>/temp/logs/ and <lab-repo>/temp/retrospectives/ to assess changes.
  • Boundary markers: There are no explicit markers or instruction-ignoring delimiters used when the agent reads this untrusted content.
  • Capability inventory: The skill possesses 'Bash' and 'Write' capabilities, allowing it to modify code and commit changes to a 'master' repository based on the ingested data.
  • Sanitization: No sanitization of the log content is performed.
  • Mitigation: The risk is significantly reduced by a mandatory human-in-the-loop (HITL) review phase (Phase 3), where the user must explicitly approve changes in a structured table before they are applied.
  • [COMMAND_EXECUTION]: Shell Command Usage. The skill utilizes the 'Bash' tool to perform repository management tasks including directory listing (ls), log inspection (git log), and source code management (git diff, git status, git commit). These commands are standard for the skill's stated purpose of backporting code changes.
Audit Metadata
Risk Level
SAFE
Analyzed
Apr 3, 2026, 06:08 PM