analyze-video-editing
Pass
Audited by Gen Agent Trust Hub on Sep 8, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [COMMAND_EXECUTION]: The scripts video_analyzer.py and audio_listener.py make extensive use of subprocess.run to interface with media processing tools including ffmpeg, ffprobe, and tesseract. These calls are implemented correctly by passing arguments as lists, which mitigates the risk of shell-based command injection.
- [INDIRECT_PROMPT_INJECTION]: The skill processes external media files (video, audio) and subtitle files (VTT). This represents an attack surface where maliciously crafted media or subtitle content could attempt to influence the agent's editorial breakdown or exploit vulnerabilities in underlying parsing libraries like librosa or opencv.
- Ingestion points: The analyze command in video_analyzer.py reads video files and VTT subtitles; audio_listener.py extracts and processes audio excerpts.
- Boundary markers: The instructions recommend reporting with specific labels (Measured, Observed, Interpreted) which helps separate machine-generated facts from interpretations, but no specific prompt boundary markers are used for subtitle text.
- Capability inventory: The skill has capabilities for file system read/write, network access to Google APIs, and execution of system binaries (ffmpeg, ffprobe, tesseract).
- Sanitization: Subtitle parsing in video_analyzer.py uses regex to strip HTML tags and timestamps, and logging uses shlex.join to safely represent commands.
- [EXTERNAL_DOWNLOADS]: The skill recommends using yt-dlp to download video content from public URLs. Additionally, audio_listener.py sends audio data to Google's official Generative Language API for semantic analysis. These operations target well-known and trusted services.
Audit Metadata