video-to-audio-foley
Pass
Audited by Gen Agent Trust Hub on Sep 20, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data, including user-supplied videos, audio references, and text prompts, creating a potential surface for indirect prompt injection.
- Ingestion points: Video URLs, text prompts, and reference audio URLs (SKILL.md).
- Boundary markers: The instructions advise using negative prompts and specific scene-level descriptions to guide the model, though they do not mandate technical delimiters for the external data itself.
- Capability inventory: The skill describes calling external APIs (fal.ai, ElevenLabs) and executing local model inference scripts (SKILL.md).
- Sanitization: The workflow relies on mandatory human-in-the-loop QA, alignment verification, and manual synchronization to identify and correct any unintended or malicious model behavior.
- [EXTERNAL_DOWNLOADS]: The skill references established research repositories and commercial AI services for audio generation.
- It directs the agent to documentation and repositories for research models like MMAudio, FoleyCrafter, and HunyuanVideo-Foley (Tencent).
- It utilizes well-known AI service providers such as fal.ai and ElevenLabs for hosted generation capabilities.
Audit Metadata