run-github-project

Pass

Audited by Gen Agent Trust Hub on Aug 28, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from GitHub issues, pull request bodies, and project comments (as seen in SKILL.md and references/normalized-ticket.md). It mitigates injection risks by instructing the agent to treat such data exclusively as 'untrusted evidence' and to verify all states against authoritative GitHub records. The capability to mutate state is protected by mandatory human-authorized Planning transitions. Data processing is handled via a validated schema in scripts/rank_tickets.py.
  • [EXTERNAL_DOWNLOADS]: The skill identifies several third-party 'provider' skills (e.g., tdd, triage, wayfinder from mattpocock/skills). While it provides npx commands for these, it enforces a critical safety policy: 'Never install a provider implicitly.' The agent is instructed to stop and notify the user if a dependency is missing, preventing unauthorized code execution.
  • [COMMAND_EXECUTION]: The skill uses the GitHub CLI (gh project) for state mutations and a local Python script (rank_tickets.py) for algorithmic ranking of the work queue. The accompanying test suite (scripts/test_rank_tickets.py) uses subprocess.run to execute the local Python script during testing, which is a standard and safe operational pattern.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 28, 2026, 03:59 AM
Security Audit — agent-trust-hub — run-github-project