The Agent Skills Directory

[COMMAND_EXECUTION]: The vault.py script executes the sops binary to decrypt local secrets. This is performed using a secure argument list to prevent shell injection and targets a specific project file.
[INDIRECT_PROMPT_INJECTION]: The skill ingests images and prompts as data for scoring by vision LLM judges.
Ingestion points: Image paths and the --prompt argument in bench.py and judge.py.
Boundary markers: The judge prompt in judge.py uses a fixed rubric template and enforces a JSON response format.
Capability inventory: Accesses external LLM APIs, writes report files, and executes the local sops binary.
Sanitization: No specific content filtering is applied to the image data or user-supplied prompts before interpolation into the judge instructions.
[SAFE]: The skill accesses the standard configuration path ~/.config/sops/age/keys.txt solely to enable SOPS decryption, representing a best practice for local secret management rather than a security risk.

vision-bench