ai-provider-claude-vision
Pass
Audited by Gen Agent Trust Hub on Oct 22, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill documents patterns for ingesting external, untrusted media files that could contain adversarial instructions. 1. Ingestion points: Patterns for reading local files and fetching content from URLs are described in SKILL.md and examples/core.md. 2. Boundary markers: Examples do not demonstrate the use of delimiters or instructions to ignore visual text. 3. Capability inventory: Uses the Anthropic SDK to interpret content. 4. Sanitization: No sanitization of visual content is demonstrated.
- [EXTERNAL_DOWNLOADS]: The skill provides examples of fetching images and PDF documents from remote URLs (e.g., example.com) for processing by the vision model.
Audit Metadata