watch-video
Pass
Audited by Gen Agent Trust Hub on Aug 14, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill utilizes yt-dlp to fetch video content and metadata from remote platforms such as YouTube, Loom, and Vimeo.\n- [COMMAND_EXECUTION]: The instructions invoke multiple shell-based utilities, including ffmpeg for frame extraction, ffprobe for duration checks, and python3 for executing the mlx-whisper transcription library.\n- [DATA_EXFILTRATION]: Video files are uploaded to the Google Gemini API (generativelanguage.googleapis.com) for multimodal analysis. As Google is a well-known service and the upload is necessary for the skill's primary function, this is considered a standard interaction.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted video transcripts and visual frame data, creating a vulnerability surface where adversarial content in the video could influence the agent's behavior.\n
- Ingestion points: Video URLs, external subtitle files, and visual frame descriptions.\n
- Boundary markers: Not present; the skill interpolates transcript text directly into synthesis prompts without explicit isolation.\n
- Capability inventory: Extensive capabilities including shell command execution (yt-dlp, ffmpeg, python3), local file writing to the user's documents, and outbound network requests.\n
- Sanitization: Includes basic filtering for VTT subtitle tags but lacks sanitization for potential instructional content within the transcript or visual analysis results.
Audit Metadata