The Agent Skills Directory

[PROMPT_INJECTION]: The skill mentions common prompt injection keywords like 'ignore' and 'instead' as part of an educational guide to help auditors detect these patterns in MCP tool definitions. It does not contain instructions attempting to override the agent's own behavioral constraints.- [COMMAND_EXECUTION]: Provides example CLI commands using node and jq for inspecting local MCP server implementations. These are standard diagnostic tools for development and auditing environments.- [EXTERNAL_DOWNLOADS]: Recommends the use of defuddle.md for fetching and converting web pages to markdown to improve token efficiency. The skill includes a 'Reading External Sources' section that explicitly cautions against routing sensitive or adversary infrastructure through this third-party service, demonstrating a strong security posture.- [DATA_EXFILTRATION]: The skill provides detailed guidance on identifying exfiltration risks in tool implementations but does not contain any code or instructions that would exfiltrate data from the user's environment.

auditing-mcp-servers