vision-language-models
Pass
Audited by Gen Agent Trust Hub on Feb 16, 2026
Risk Level: LOW
Full Analysis
- [PROMPT_INJECTION] (LOW): The skill documents patterns for processing external images and PDFs, which inherently presents an indirect prompt injection surface. * Ingestion points: File paths for PDFs and images are processed in references/document-vision.md. * Boundary markers: The provided snippets do not include explicit delimiters or instructions to ignore embedded text directives within the documents. * Capability inventory: Data is passed to high-reasoning models (Claude, Gemini) for extraction and analysis. * Sanitization: No sanitization logic is included in the reference patterns.
- [SAFE] (SAFE): Analysis of the provided Python snippets for image resizing, batch processing, and model tiering confirms they are functional references for API integration without any hidden malicious payloads or suspicious network activity.
Audit Metadata