segment-anything-model

Pass

Audited by Gen Agent Trust Hub on Sep 9, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: Fetches model weights and essential dependencies from Meta AI's official public hosting and GitHub repositories. This includes Segment Anything checkpoints from dl.fbaipublicfiles.com and source code from the facebookresearch organization on GitHub.
  • [INDIRECT_PROMPT_INJECTION]: The skill documents how to process images and user-supplied prompts (points, boxes), creating a data ingestion surface for the underlying vision models.
  • Ingestion points: Processes images via cv2.imread and Image.open, and accepts coordinate-based inputs in SKILL.md and references/advanced-usage.md.
  • Boundary markers: The code examples provided do not include specific instructions or markers to the model to disregard potential adversarial content contained within the input data.
  • Capability inventory: The documented capabilities are restricted to computer vision tasks (mask generation, object segmentation); no risky capabilities such as arbitrary command execution, network exfiltration of processed data, or file writing are enabled in the provided code.
  • Sanitization: The implementation guides use standard library calls and do not demonstrate additional layers of input validation or content filtering for incoming images or prompt arrays.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 9, 2026, 07:07 PM
Security Audit — agent-trust-hub — segment-anything-model