The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes untrusted source code from local repositories, creating a surface for malicious instructions to influence the agent or downstream tools. * Ingestion points: Local repository files and .dev/codepatrol/ configuration files (e.g., targets.md, checklist.md). * Boundary markers: The skill lacks explicit delimiters or instructions to treat potentially malicious code as data rather than instructions when passing summarized findings to the codex-consultation tool. * Capability inventory: File system read/write access, interaction with the codex-consultation tool, and writing to the cosense platform. * Sanitization: No sanitization or escaping of the processed code is performed before inclusion in analysis prompts.
[DYNAMIC_EXECUTION]: The instructions explicitly guide the agent and downstream tools to execute code found within the untrusted target repository for verification purposes. * Evidence: Section 手順2 (Step 2) includes instructions to run code in a temporary directory outside the repository to verify security findings, which involves executing arbitrary code found in the audited files.
[PROMPT_INJECTION]: The skill contains specific strategies designed to evade the safety and content filters of the downstream codex-consultation (Codex) tool. * Evidence: The 依頼の枠組み (Request Framework) section instructs the agent to avoid using terms like "attack", "exploit", or "forgery" and to frame requests as "fixing our own code" to prevent the model provider's security filters from blocking analysis.
[COMMAND_EXECUTION]: The skill instructs the agent to perform extensive file system operations and tool-based executions based on external, unverified project configurations.

codepatrol-report