lookdev-auto
Pass
Audited by Gen Agent Trust Hub on Sep 4, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONNO_CODE
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill instructions involve processing output from external AI models (vision/video judges). This creates a potential surface for indirect prompt injection if the model output is untrusted. \n
- Ingestion points: Output from external vision/video models containing ratings and parameter suggestions.\n
- Boundary markers: Prompting the model to return "ONLY JSON" and advising the agent to regex the first
{...}block.\n - Capability inventory: The agent is expected to use rendering tools and write parameters to files.\n
- Sanitization: Recommends regex-based parsing to isolate structured data from potential prose.\n- [NO_CODE]: The skill contains only documentation, metadata, and instructions. No executable code or scripts are provided.\n- [SAFE]: External service references, such as those to
fal-ai, and external links to the author's GitHub repositories and badge providers, target well-known and legitimate services.
Audit Metadata