rebut

Warn

Audited by Gen Agent Trust Hub on Jun 20, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill executes multiple shell scripts (preflight.sh, gather-review.sh, and ci-status.sh) by constructing paths using the ${CLAUDE_PLUGIN_ROOT} environment variable. This dynamic path resolution for executable scripts is a noted pattern for runtime execution.
  • [PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface because it ingests untrusted content from pull request comments and possesses capabilities to modify the environment and repository.
  • Ingestion points: Gathers all PR review summaries and inline comments via gather-review.sh in Step 1.
  • Boundary markers: Utilizes 'HARD-GATE' sections that explicitly instruct the agent to never apply or refute suggestions without verification against current code and repository gates.
  • Capability inventory: The agent can read files (Step 2), execute local commands such as lint/tests (Step 2), modify source code and commit/push changes via git (Step 4), and post new comments to the GitHub API (Step 5 and 6).
  • Sanitization: Relies on the mandatory manual verification process (reproducing the claim and running repo gates) to sanitize the impact of potentially malicious suggestions embedded in comments.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Jun 20, 2026, 02:28 AM
Security Audit — agent-trust-hub — rebut