watch-video

Pass

Audited by Gen Agent Trust Hub on Nov 17, 2026

Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill frequently uses subprocess.run and subprocess.Popen in scripts such as watch.py, perceive.py, and read_screen.py to invoke system tools like ffmpeg, ffprobe, and yt-dlp. These executions are essential for media processing and are performed using structured arguments to minimize injection risks.
  • [EXTERNAL_DOWNLOADS]: Model weights are downloaded at runtime from external repositories. scripts/diarize.py fetches ONNX models from GitHub releases (k2-fsa/sherpa-onnx), and scripts/transcribe_qwen.py utilizes snapshot_download from Hugging Face. Both are well-known and established sources for AI research and tooling.
  • [INDIRECT_PROMPT_INJECTION]: The skill has a significant attack surface for indirect prompt injection as it ingests large amounts of untrusted text from video content.
  • Ingestion points: Untrusted data enters the agent context through video metadata, comments, transcripts (processed in SKILL.md), and OCR text (generated in read_screen.py).
  • Boundary markers: Present. The instructions explicitly state that content from the video must be treated as data and never as instructions.
  • Capability inventory: The skill can execute arbitrary shell commands via multiple scripts and perform file system operations.
  • Sanitization: Absent. There is no evidence of automated sanitization or escaping for the ingested text.
  • [DYNAMIC_EXECUTION]: A static analysis alert for eval() in scripts/read_screen.py was identified as a false positive. The script calls the .eval() method on a PyTorch model object to set it to evaluation mode, which is a standard and secure practice in machine learning workflows and does not execute arbitrary code strings.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 17, 2026, 11:51 AM