video-delivery-coach
Pass
Audited by Gen Agent Trust Hub on Nov 6, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONMETADATA_POISONING
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted video content and interpolates the resulting transcription directly into the LLM prompt without sufficient isolation.
- Ingestion points: Video transcriptions are extracted from user-provided files in
scripts/analyze_video.py. - Boundary markers: The prompt used in
generate_coaching_reportlacks specific delimiters (such as XML tags or triple backticks) or instructions to ignore commands embedded within the transcription text. - Capability inventory: The skill possesses file read/write access and the ability to perform network requests to the Anthropic API.
- Sanitization: There is no sanitization or filtering of the transcribed content before it is processed by the model.
- [METADATA_POISONING]: The skill makes misleading claims about its privacy and execution model in the documentation.
- Evidence: The
SKILL.mdfile states that "All analysis is local, video never uploaded anywhere." While the raw video file itself is not uploaded,scripts/analyze_video.pysends the full text transcription and extracted facial metrics to an external API (Anthropic) to generate the coaching report. This contradicts the claim of entirely local analysis.
Audit Metadata