media-judge
Pass
Audited by Gen Agent Trust Hub on Nov 4, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: Fetches pre-trained OpenCV face detection and recognition models (YuNet and SFace) from the official Hugging Face repository. This is used for local identity and continuity analysis during the media review process.
- [COMMAND_EXECUTION]: Extensively uses the system's FFmpeg and ffprobe binaries via subprocess calls to perform frame extraction, audio analysis, and compression simulation.
- [INDIRECT_PROMPT_INJECTION]: Processes untrusted content such as on-screen text (via OCR) and speech-to-text transcripts. This constitutes an indirect prompt injection surface. The skill mitigates this by providing a structured 'vision pass' protocol (defined in references/eyes.md and references/observe-prompt.txt) that requires the agent to describe visual elements and quote specific frames before making judgments, reducing the likelihood of the agent obeying instructions embedded in the media.
- [DYNAMIC_EXECUTION]: Generates and executes dynamic FFmpeg filter chains and Python analysis scripts based on user-provided metadata and configuration flags such as safe zone detection and bitrate simulation.
Audit Metadata