lookdev-auto
Pass
Audited by Gen Agent Trust Hub on Oct 27, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill defines a feedback loop that processes model-generated interpretations of visual data.
- Ingestion points: Renders (images/videos) are processed by vision models to extract tuning parameters as described in the iteration process.
- Boundary markers: The instructions recommend using "explicit rubrics" and "structured JSON output" to constrain the model's judgment and prevent misinterpretation.
- Capability inventory: The skill utilizes external CLI tools (claude-code, gemini-cli, etc.) to perform rendering and API-based visual inference.
- Sanitization: The skill advises using regex-based parsing of JSON output to isolate relevant data from model responses.
- [EXTERNAL_DOWNLOADS]: The skill documentation refers to the use of external cloud-based AI services for visual analysis.
- Evidence: References utilizing fal-ai/video-understanding for temporal and motion-based evaluation.
Audit Metadata