babysit-pr

Pass

Audited by Gen Agent Trust Hub on Aug 22, 2026

Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [PROMPT_INJECTION]: The skill processes data from pull request threads and reviews which are external, untrusted sources. This creates a surface for indirect prompt injection where malicious PR comments could influence the agent's behavior during its autonomous fixing loop.\n
  • Ingestion points: Pull request thread bodies and review summaries fetched via the scripts/threads.sh script from GitHub or GitLab.\n
  • Boundary markers: The skill uses metadata markers like <!-- debate-review:... --> to identify specific bot comments, but it lacks robust isolation or escaping for general review content which is parsed and followed by the agent.\n
  • Capability inventory: The skill can perform repository writes via git push, interact with forge APIs (GitHub/GitLab) to resolve threads and post replies, and execute local scripts as well as repository-defined build/test gates.\n
  • Sanitization: No explicit sanitization or validation of the external review content is performed before the agent uses it to decide on and implement code fixes.\n- [COMMAND_EXECUTION]: The skill executes various system tools and scripts to manage the PR lifecycle.\n
  • Evidence: Execution of scripts/threads.sh to harvest PR data and usage of gh api and glab api for pull request interactions.\n
  • Evidence: The skill instructs the agent to run the repository's own test suite (the repo's own gate) and external skill scripts such as debate-review/scripts/review-pr.mjs.\n
  • Evidence: The bash harvesting script uses standard command-line utilities (git, sed, jq) to process forge metadata and git origin URLs.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 22, 2026, 12:11 AM
Security Audit — agent-trust-hub — babysit-pr