segment-anything-model

Pass

Audited by Gen Agent Trust Hub on Sep 8, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill instructs the user to install the segment-anything and segment-anything-2 packages directly from official GitHub repositories managed by Facebook Research. It also provides links to download model checkpoints (ViT-H, ViT-L, ViT-B) from Meta's authenticated public file servers. These are recognized official sources for the research model weights.
  • [INDIRECT_PROMPT_INJECTION]: The advanced usage section includes an implementation of 'Grounded SAM' which processes natural language text prompts to identify and segment objects in images. This creates a surface for indirect prompt injection if the text prompts are derived from untrusted external sources.
  • Ingestion points: The text_prompt parameter in the text_to_mask function in references/advanced-usage.md.
  • Boundary markers: Absent; the text prompt is passed directly to the model's inference function without delimiters.
  • Capability inventory: The skill uses model inference to generate image segmentations based on text prompts.
  • Sanitization: Absent; the provided code does not demonstrate validation or filtering of the text input strings.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 8, 2026, 07:01 PM
Security Audit — agent-trust-hub — segment-anything-model