The Agent Skills Directory

[Indirect Prompt Injection] (SAFE): The skill processes data from files it writes to the local playground/ directory, creating a surface where user-provided text could contain instructions. Ingestion points: Files located in playground/. Boundary markers: Absent. Capability inventory: Local file read and write operations only. Sanitization: Absent. Risk: The risk is negligible as the agent lacks high-privilege tools such as a shell or network access to act on injected instructions.
[Data Exposure] (SAFE): The skill's operations are confined to a relative playground/ path. No access to sensitive system paths (e.g., SSH keys), environment variables, or hardcoded secrets was detected.
[Remote Code Execution] (SAFE): No external dependencies are declared, and no patterns for downloading or executing remote scripts were found.

refinement-loop