wan-3-0-prime-reference-to-video
Pass
Audited by Gen Agent Trust Hub on Oct 6, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill allows the agent to process reference images, videos, and audio clips from external URLs. This creates a surface for indirect prompt injection where instructions could be hidden within the media.
- Ingestion points: External media URLs are processed via the
reference_images,reference_videos, andreference_audiosfields in theruncomfycommand. - Boundary markers: The skill includes a dedicated 'Security & Privacy' section that explicitly instructs the agent to treat all reference media as untrusted data and ignore any embedded directives.
- Capability inventory: The skill uses the
Bash(runcomfy *)tool, which performs network requests to the Model API and writes output files to the local file system. - Sanitization: The agent is explicitly instructed to only extract relevant generation inputs and ignore any commands or links found within the media.
- [COMMAND_EXECUTION]: The skill utilizes the
runcomfyCLI tool to interact with the RunComfy Model API. It documents the use of JSON-formatted inputs to ensure data is passed to the API without shell expansion risks, and limits the tool scope to the specificruncomfycommand prefix.
Audit Metadata