video-understand
Pass
Audited by Gen Agent Trust Hub on Nov 14, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The script
understand_video.pymakes extensive use of thesubprocessmodule to callffmpeg,ffprobe, and thewhisperCLI. These calls are implemented securely by passing arguments as lists rather than raw strings, which mitigates the risk of shell command injection. - [EXTERNAL_DOWNLOADS]: The skill includes an auto-install feature for the
openai-whisperpackage. If the package is missing, the script will prompt the user in interactive sessions to install it viapipfrom the official Python Package Index (PyPI). This is a well-known package from a reputable source, and the installation requires user consent. - [INDIRECT_PROMPT_INJECTION]: The skill transcribes audio from user-provided video files and returns the resulting text to the agent's context. This creates a potential vector for indirect prompt injection attacks where malicious instructions embedded in a video's audio could influence the agent's behavior.
- Ingestion points: Audio data from video files is processed and transcribed in
scripts/understand_video.py. - Boundary markers: The transcript text is returned as a plain string in the JSON response without explicit delimiters or instructions for the agent to ignore commands within the content.
- Capability inventory: The skill can execute local media processing commands and perform file writes to a specific output directory.
- Sanitization: The transcribed text is not sanitized or filtered before being returned to the agent context.
Audit Metadata