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.com and HuggingFace.
  • [REMOTE_CODE_EXECUTION]: Instructions guide the user to install libraries directly from Meta Research's GitHub repositories (facebookresearch/segment-anything and facebookresearch/segment-anything-2) using pip install git+....
  • [COMMAND_EXECUTION]: Provides standard shell commands for environment setup, including package installation via pip and file retrieval via wget.
  • [INDIRECT_PROMPT_INJECTION]:
  • Ingestion points: The skill processes external image files provided by users via cv2.imread, FastAPI UploadFile endpoints, 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
Risk Level
SAFE
Analyzed
Aug 28, 2026, 12:47 PM
Security Audit — agent-trust-hub — segment-anything-model