gemini-video-understanding
Pass
Audited by Gen Agent Trust Hub on Sep 4, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes external video content and user-provided prompts via Google's Gemini API, creating a vector for instructions embedded within data to influence the agent's behavior.
- Ingestion points: Video files are ingested via
scripts/analyze_video.pyfrom local paths or YouTube URLs. User-supplied text is passed as a prompt to the model. - Boundary markers: The script does not implement explicit boundary markers or instruction isolation when interpolating the user prompt with video data in the API request.
- Capability inventory: The skill uses the
google-genailibrary to interact with Google's API. It has file read access (to upload videos) and shell execution capabilities through the Claude agent. - Sanitization: No specific sanitization or filtering is performed on the content of the video or the prompt before transmission to the Gemini API.
Audit Metadata