video-transcript

Pass

Audited by Gen Agent Trust Hub on Nov 12, 2026

Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill uses node to execute internal scripts (doctor.mjs, transcribe.mjs) which accept $ARGUMENTS directly from user input. While the scripts appear to be wrappers for a shared engine, there is a risk of shell argument injection if the underlying implementation doesn't strictly sanitize inputs.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data (video transcripts from YouTube, Bilibili, and Douyin).
  • Ingestion points: Video metadata, subtitles, and Whisper-generated transcripts from external URLs or files.
  • Boundary markers: Absent. The skill lacks explicit delimiters or warnings to the model to ignore instructions embedded within the transcripts themselves.
  • Capability inventory: File system read/write, network operations (via yt-dlp), and shell command execution.
  • Sanitization: The skill performs text cleaning (OpenCC conversion, punctuation fixes) but does not specifically sanitize for prompt injection patterns in the processed text.
  • [DATA_EXPOSURE]: The skill mentions using --cookies-from-browser chrome to bypass bot detection. However, it explicitly instructs the agent to ask for user consent first and forbids the scripts from automatically reading cookies or installing external dependencies, which is a positive safety control.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 12, 2026, 04:48 PM