The Agent Skills Directory

[PROMPT_INJECTION]: Documentation files such as references/ai-agent-security.md and references/owasp-checklists.md contain strings that trigger prompt injection detectors, including 'Ignore all previous instructions' and 'DAN' jailbreak patterns. These are used as educational examples for security testing. The skill provides clear instructions for the agent to treat these strings as test data rather than instructions, significantly reducing the risk of the agent adopting a malicious role.
[INDIRECT_PROMPT_INJECTION]: The skill analyzes external code and configuration files, creating a potential surface for indirect prompt injection.
Ingestion points: Files in the target workspace processed by full_audit.py, quick_scan.py, and the specialized scanners in the scanners/ directory.
Boundary markers: Present. The skill includes a 'Review boundary' disclaimer at the top of all major reference and skill files to distinguish between exploit examples and operational instructions.
Capability inventory: The skill scripts possess local file system read access for analysis and write access for logging and report generation. There are no network capabilities or subprocess execution tools provided within the skill's own code.
Sanitization: The scanning logic uses regex for static pattern matching. While it does not sanitize the content of analyzed files, the risk is mitigated because the content is handled as data for reporting rather than being interpolated into executable logic.

007