The Agent Skills Directory

[COMMAND_EXECUTION]: The skill explicitly instructs the agent to execute code on the host system to verify safety claims. It directs the agent to "prove them by running code" and specifically to "Write a script or test that runs the real code, run it, and paste what happened."
[DYNAMIC_EXECUTION]: The agent is encouraged to dynamically generate new scripts or tests based on the code it is analyzing to confirm how the software behaves in practice.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input in the form of code changes, diffs, and pull request data which could theoretically contain malicious instructions.
Ingestion points: Analyzes diffs, symbols, PRs, and commits (SKILL.md).
Boundary markers: No specific delimiters are required for the untrusted code data.
Capability inventory: Execution of scripts and tests on the local machine (SKILL.md).
Sanitization: The skill advises the agent to "strip anything private before it goes anywhere public," which provides some data protection, though it does not explicitly sanitize input data to prevent injection.
[EXTERNAL_DOWNLOADS]: The skill metadata in MIRROR.md references the official GitHub repository for Cursor plugins, a well-known development service.

cursor-blast-radius