code-review

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 0.80). The skill ingests untrusted, user-contributed code diffs and files and passes them directly to external LLM CLIs in its required workflows — e.g., the GitHub Actions (gemini-review.yml / codex-review.yml / dual/triple-review workflows) explicitly run git diff origin/${{ github.base_ref }}...HEAD and feed that diff to gemini/codex/claude, and the pre-commit hook invokes claude on staged files — meaning arbitrary third-party PR content is read and can influence the agents' decisions.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.90). The skill contains runtime steps that fetch and execute remote code — e.g., the GitHub Action reference uses: openai/codex-action@main, the npx invocation of @anthropic-ai/claude-code, and the Gemini extension install URL https://github.com/gemini-cli-extensions/code-review — all of which are pulled/installed at runtime and can execute remote code or control agent prompts/instructions.