babysit

Warn

Audited by Gen Agent Trust Hub on Nov 8, 2026

Risk Level: MEDIUMINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from GitHub Pull Request comments and review threads which could contain malicious instructions.\n
  • Ingestion points: Untrusted content is retrieved from external sources using gh pr view and get-pr-comments in SKILL.md.\n
  • Boundary markers: The skill lacks explicit delimiters or "ignore" instructions when interpolating this external data into its internal triage and action logic.\n
  • Capability inventory: The skill possesses capabilities to modify the local codebase (via fix-ci and conflict resolution) and execute system commands.\n
  • Sanitization: There is no evidence of filtering or sanitizing the content of PR comments before the agent triages them or replies to them.\n- [DYNAMIC_EXECUTION]: The skill dynamically generates and executes JavaScript code via a shell command, creating a risk of code injection.\n
  • Evidence: In SKILL.md (Step 4), a node -e command is used to record entries in a verification ledger. The script is built by interpolating variables like <N>, <HEAD_SHA>, and <command> directly into a JavaScript string literal.\n
  • Risk: If any of these variables (particularly <command>, which may come from external check names or tool output) contain single quotes or malicious JavaScript fragments, it could break the string literal and execute arbitrary code.\n
  • Implementation details: The script uses require.resolve and dynamic import() with paths computed from environment variables (CLAUDE_PLUGIN_ROOT, OMC_PLUGIN_ROOT) and local package searches.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 8, 2026, 07:22 PM