youtube-harvester
Pass
Audited by Gen Agent Trust Hub on Sep 20, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill uses Python's subprocess.run to call yt-dlp for metadata extraction and playlist processing. Evidence: subprocess.run calls in get_video_metadata, harvest_youtube_playlist, and harvest_youtube_channel in SKILL.md.
- [EXTERNAL_DOWNLOADS]: The skill fetches content from YouTube and downloads semantic embedding models from Hugging Face for topic chunking. Evidence: YouTubeTranscriptApi and yt-dlp usage for data retrieval, and SentenceTransformer initialization for model loading.
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted transcript data from YouTube which represents a potential surface for indirect prompt injection. 1. Ingestion points: The harvest_youtube_video function in SKILL.md. 2. Boundary markers: Not explicitly defined in the provided snippets. 3. Capability inventory: Subprocess execution and RAG ingestion. 4. Sanitization: No explicit sanitization of transcript content is performed.
Audit Metadata