segment-anything-model
Pass
Audited by Gen Agent Trust Hub on Aug 28, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill fetches large pre-trained model checkpoints from Meta AI's official research storage at
dl.fbaipublicfiles.comand HuggingFace. - [REMOTE_CODE_EXECUTION]: Instructions guide the user to install libraries directly from Meta Research's GitHub repositories (
facebookresearch/segment-anythingandfacebookresearch/segment-anything-2) usingpip install git+.... - [COMMAND_EXECUTION]: Provides standard shell commands for environment setup, including package installation via
pipand file retrieval viawget. - [INDIRECT_PROMPT_INJECTION]:
- Ingestion points: The skill processes external image files provided by users via
cv2.imread, FastAPIUploadFileendpoints, and Gradio image inputs. - Boundary markers: None identified for image data processing.
- Capability inventory: The skill includes file system access for loading models and saving images, network operations for model downloads, and external package management.
- Sanitization: Standard image decoding via OpenCV and PIL is used without explicit input validation beyond format conversion.
Audit Metadata