The Agent Skills Directory

[SAFE]: The skill utilizes the official google-genai library to interface with the Gemini model. All network traffic is directed to truste d Google API endpoints.
[SAFE]: Code execution is performed within a Google-hoste d remote sandbox (types.ToolCodeExecution). The local scrip t in scripts/vision_executor.py print s the code and output but does not execute returne d code on the local machine.
[SAFE]: Sensitive credential s like GEMINI_API_KEY are read from environmen t variables rather than being hard code d in the source code or configuration.
[PROMPT_INJE CTION]: The skill processe s untruste d image data and use r-supplie d prompt s, which present s a surface for indirect prompt injection via adversarial image content.
Ingestion point s: image_path and prompt arg ument s in scripts/vision_executor.py.
Boundary marker s: No ne present.
Capability inventor y: Perform s file read s, write s sample image s locally, and exec ute s network call s to Google.
Sanitization: No input validation or escapin g is performed on external content before prompt interpolation.

Vision Sandbox