The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes documentation on LLM security (e.g., in domains/ai/llm-security.md) that lists prompt injection and jailbreak techniques like 'DAN'. These references are strictly for educational purposes and defensive guidance, not active attempts to subvert the agent.
[COMMAND_EXECUTION]: The skill contains several Node.js scripts (quality_checker.js, module_scanner.js, change_analyzer.js, security_scanner.js, doc_generator.js) that use standard system utilities (like git) and file operations to audit projects and generate reports. These actions are consistent with the skill's primary purpose.
[INDIRECT_PROMPT_INJECTION]: The auditing and documentation tools ingest untrusted project data, creating a potential surface for indirect prompt injection.
Ingestion points: Scripts such as doc_generator.js and security_scanner.js read local project files.
Boundary markers: The generated reports and README files do not use explicit boundary markers to isolate user-supplied content.
Capability inventory: The agent can read/write files and execute shell commands.
Sanitization: The tool focus is on regex analysis rather than input sanitization for the generated output.
Risk: Although a surface exists, it is inherent to the tool's auditing function, and the overall risk is minimal due to the localized nature of the operations.

sage