The Agent Skills Directory

Command Execution: The skill invokes internal scripts like generate_rank_input.py and launch_codex_security_mcp to manage the discovery and ranking of security findings. These tools are executed locally as part of the scan workflow.
Indirect Prompt Injection Surface: The skill is designed to analyze external code changes, which involves processing untrusted data.
Ingestion points: The tool reads repository files and diff inventories to find candidate security issues (SKILL.md).
Boundary markers: The skill does not specify the use of delimiters or 'ignore' instructions when reading untrusted repository content.
Capability inventory: The agent can access files and execute local helper scripts via Python to facilitate the security review process.
Sanitization: The instructions do not define specific validation or sanitization steps for the code being scanned before it is processed by helper tools.

finding-discovery