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.comand source code from thefacebookresearchorganization 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.imreadandImage.open, and accepts coordinate-based inputs inSKILL.mdandreferences/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