The Agent Skills Directory

Indirect Prompt Injection (HIGH): The skill is highly vulnerable to instructions embedded within the code it analyzes. Malicious comments in a processed code snippet could override the agent's logic or force it to report a false 'SAFE' status.
Evidence Chain:
Ingestion points: Tools analyze_code_with_codex, analyze_code_with_gemini, and analyze_code_combined take raw code as input.
Boundary markers: Absent. There is no evidence of delimiters or 'ignore' instructions for the external content.
Capability inventory: The skill produces 'Full Audit' reports and 'Security Scan' results which influence user trust and merging decisions.
Sanitization: No sanitization of the input code is described.
Data Exfiltration & Exposure (MEDIUM): The skill's primary function is sending local code snippets to external AI models (OpenAI/Google). This constitutes a data exposure risk for proprietary or sensitive source code.
External Downloads & Dependencies (LOW): The skill requires pre-installed 'Codex CLI' and 'Gemini CLI'. While Google-Gemini is a trusted source, the specific provenance of the 'Codex CLI' is not provided, posing a potential supply chain risk.
No Code Provided (INFO): The submission contains only documentation (SKILL.md) and no executable script files. The actual behavior of the MCP tools and their handling of data cannot be verified.

ai-code-reviewer