The Agent Skills Directory

[UNVERIFIABLE_DEPENDENCIES] (MEDIUM): The file scripts/download_video.py contains logic in check_yt_dlp() to automatically install the yt-dlp package from PyPI using pip at runtime. Dynamic package installation can be exploited if an attacker intercepts the connection or if the package registry is compromised.
[INDIRECT_PROMPT_INJECTION] (LOW): The skill ingests untrusted data from an external source (YouTube video titles and metadata via yt-dlp --dump-json).
Ingestion points: Video title and uploader name are fetched from the YouTube API.
Boundary markers: None. The metadata is printed directly to the console.
Capability inventory: The script has filesystem write access and network access.
Sanitization: None. Malicious instructions embedded in a video title could potentially influence the agent's next steps if it reads the console output.
[COMMAND_EXECUTION] (SAFE): The script uses subprocess.run() with a list of arguments (e.g., cmd = ["yt-dlp", ... url]) instead of a single string with shell=True. This is a best practice that prevents shell injection vulnerabilities.

youtube-downloader