skills/ikatkov/agent-skills/review-pr/Gen Agent Trust Hub

review-pr

Pass

Audited by Gen Agent Trust Hub on Aug 10, 2026

Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill relies on system commands including 'git', 'gh', 'jq', and 'curl' to manage repository state and query the GitHub API. It includes a functional routine to copy the internal helper script 'wait-for-reviews.sh' to a temporary directory ('$TMPDIR') for execution, which is described as a workaround for sandbox environments that restrict reading from symlinked skill paths.- [PROMPT_INJECTION]: The skill processes external PR data and review feedback, presenting an indirect prompt injection surface. This is mitigated through explicit instructions for the agent to triage all findings against a 'SCOPE' contract defined by the PR title and description before making code changes.
  • Ingestion points: PR comments, review bodies, and threads fetched via the GitHub REST/GraphQL APIs (SKILL.md, Section 3).
  • Boundary markers: The 'SCOPE' contract (defined in Section 1 and enforced in Section 4) serves as a logical boundary for acceptable modifications.
  • Capability inventory: File system write operations, 'git commit', 'git push', and GitHub API interactions.
  • Sanitization: Relies on agent-led triage and verification of required checks as specified in the skill instructions.- [SAFE]: Authentication is handled using standard environment variables ('GH_TOKEN') to communicate with official GitHub API endpoints. No obfuscation, persistence mechanisms, or exfiltration of sensitive local data (e.g., SSH keys or AWS credentials) were observed.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 10, 2026, 10:54 PM
Security Audit — agent-trust-hub — review-pr