The Agent Skills Directory

[PROMPT_INJECTION]: The skill possesses a vulnerability surface for indirect prompt injection due to its recursive processing of untrusted repository data while maintaining high-privilege access.\n
Ingestion points: Ingests content from multiple locations including .claude/hooks/*.py, .claude/hooks/*.sh, .claude/skills/*/SKILL.md, and project documentation files.\n
Boundary markers: The workflow lacks specific boundary markers or instructions to isolate the audited content from the agent's logic, which could lead to the execution of instructions embedded within the files.\n
Capability inventory: The skill is equipped with Write, Edit, Bash, and Task tools, allowing it to modify code and execute system-level commands based on the findings.\n
Sanitization: There is no evidence of sanitization or strict schema validation for the data ingested from the filesystem before it triggers fixes or command execution.\n- [COMMAND_EXECUTION]: The skill performs shell operations and interacts with executable files as part of its auditing and build processes.\n
Evidence: Utilizes the Bash tool to perform documentation rendering tasks such as quarto render.\n
Evidence: Explicitly identifies and modifies logic within executable hook scripts in the .claude/hooks/ directory to enforce code quality standards.

deep-audit