The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the user to download model weights and libraries from established external sources.
Fetches the segment-anything and segment-anything-2 libraries directly from Facebook Research's official GitHub repository.
Downloads pre-trained model checkpoints (ViT-H, ViT-L, ViT-B) from dl.fbaipublicfiles.com, which is a well-known service operated by Meta AI.
These operations target well-known organizations and do not involve piped execution into a shell.
[COMMAND_EXECUTION]: The documentation includes standard setup commands for a development environment.
Provides pip install commands for necessary Python dependencies like transformers, torch, opencv-python, and pycocotools.
Provides wget commands to retrieve large model files required for the skill's primary function.
Includes an ONNX export script example for model deployment.
[INDIRECT_PROMPT_INJECTION]: The skill defines an interface that processes external data, creating a potential attack surface.
Ingestion points: Processes images via cv2.imread, FastAPI's UploadFile, and Gradio's gr.Image. Also accepts point and box coordinate prompts from users.
Boundary markers: Code examples do not explicitly demonstrate sanitization of image metadata or validation of coordinate bounds beyond basic assertions in the troubleshooting guide.
Capability inventory: The skill facilitates local file writes (saving masks/annotations) and model downloads.
Sanitization: Standard for this type of vision task, the vulnerability is inherent to processing complex binary data (images).

segment-anything-model