The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes the Bash and Write tools for codebase research but implements robust security constraints, including mandatory validation of the --scope argument to reject absolute paths and directory traversal (..) attempts.
[INDIRECT_PROMPT_INJECTION]: The skill handles untrusted data fetched from the web through WebFetch and WebSearch tools. It implements a clear defense-in-depth strategy by instructing the agent to ignore instructions within fetched pages, cross-verify all claims, and never execute code found in external sources.
[PROMPT_INJECTION]: The skill proactively identifies user-provided inputs such as the research <topic> as untrusted and explicitly forbids their interpolation as executable instructions, protecting against direct prompt injection.

deep-research