qwen-mm-plugins-video-spatio
Pass
Audited by Gen Agent Trust Hub on Sep 27, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data in the form of video frames and images to ground objects and reason about their spatial relationships. This represents a potential surface for indirect prompt injection if the visual data contains adversarial content designed to influence the vision-language model (VLM).
- Ingestion points: Video and image data are ingested via the core tool's read_video function or by providing local frame paths to the build_scene tool.
- Boundary markers: The skill does not explicitly specify boundary markers or delimiters for raw visual data processing, relying on the model's internal grounding and labeling logic.
- Capability inventory: The skill utilizes tools for file writing (save_view) and network-based VLM processing via OpenAI-compatible endpoints.
- Sanitization: No explicit sanitization or filtering of image/video content is described; the workflow relies on visual grounding and cross-checks like verify_grounding.
- [SAFE]: The skill focuses on geometric calculations and spatial reasoning instructions. It uses stateless tools for scene construction and visualization, which trace back to the primary intended functionality without signs of malicious behavior or obfuscation.
Audit Metadata