raw-to-reel

Pass

Audited by Gen Agent Trust Hub on Aug 21, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill fetches machine learning models for audio transcription from the Hugging Face model registry.\n
  • Evidence: The Python transcription snippet references the mlx-community/whisper-small.en-mlx repository.\n- [COMMAND_EXECUTION]: The skill drives video processing through the FFmpeg suite and utilizes Python's subprocess module for metadata extraction.\n
  • Evidence: Instructions provide detailed ffmpeg filter graphs and a script that executes ffprobe via subprocess.run to calculate clip durations.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external video footage and generated transcripts. While these represent a potential attack surface for embedded instructions, the skill limits exposure by using the data strictly for technical alignment and timestamp calculation.\n
  • Ingestion points: Raw video files (e.g., .mp4, .mov) and Whisper-derived word-level transcripts.\n
  • Boundary markers: Not explicitly used; the workflow relies on visual and technical verification of the output.\n
  • Capability inventory: Includes full access to ffmpeg, ffprobe, and shell command execution for rendering.\n
  • Sanitization: The skill emphasizes manual spot-checks and automated integrity verification (e.g., ffprobe stream checks) to identify processing anomalies.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 21, 2026, 12:07 PM
Security Audit — agent-trust-hub — raw-to-reel