video-summary

Pass

Audited by Gen Agent Trust Hub on Sep 19, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes external, untrusted content (video transcripts and subtitles) which is then interpolated into a structured request for an LLM summary. This creates a surface for indirect prompt injection if the video content contains malicious instructions intended to manipulate the LLM's output.
  • Ingestion points: Transcripts are fetched via yt-dlp or generated locally using openai-whisper within the video-summary.sh script.
  • Boundary markers: The call_llm function in video-summary.sh incorporates the transcript text into a summary prompt without utilizing distinct delimiters or instructions for the LLM to ignore embedded commands within the transcript.
  • Capability inventory: The script performs file system reads (for cookies and local videos), writes temporary files to /tmp, and executes several CLI tools including yt-dlp, ffmpeg, and whisper.
  • Sanitization: A clean_subtitle function is present in video-summary.sh to remove technical subtitle formatting (like WebVTT/SRT tags), though it does not filter the semantic content of the speech.
  • [COMMAND_EXECUTION]: The main execution script video-summary.sh orchestrates several third-party binaries (yt-dlp, ffmpeg, jq, whisper) via shell subprocesses. While arguments are generally quoted to prevent simple command injection, the skill's safety relies on the robustness of these external tools when processing potentially malformed video metadata or content from remote URLs.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 19, 2026, 11:32 PM
Security Audit — agent-trust-hub — video-summary