video-to-text

Warn

Audited by Gen Agent Trust Hub on Mar 7, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCREDENTIALS_UNSAFEPROMPT_INJECTION
Full Analysis
  • [REMOTE_CODE_EXECUTION]: Executes a shell script (download-ggml-model.sh) from a user-provided or repository-relative path. Evidence: bash download-ggml-model.sh base within the project directory.
  • [COMMAND_EXECUTION]: Provides instructions to run various system-level commands and start a local web server, increasing the local attack surface. Evidence: npm link, brew install, video2text serve, video2text extract.
  • [EXTERNAL_DOWNLOADS]: Fetches AI models and installs packages from the NPM registry and third-party Homebrew taps. Evidence: npm install -g @wangjs-jacky/video2text, brew tap fyrfyrr/f2.
  • [CREDENTIALS_UNSAFE]: Encourages the use of raw authentication cookies as command-line arguments, which can lead to sensitive data exposure in process lists or logs. Evidence: video2text extract <URL> -c "<cookie内容>".
  • [PROMPT_INJECTION]: Vulnerable to indirect prompt injection via malicious video URLs or platform metadata. Evidence: 1. Ingestion: Video URLs and platform metadata via yt-dlp. 2. Boundary markers: Absent. 3. Capabilities: Subprocess execution (ffmpeg, yt-dlp), network access, and local file writes. 4. Sanitization: No explicit sanitization of external metadata.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Mar 7, 2026, 04:53 PM