The Agent Skills Directory

[Unverifiable Dependencies] (MEDIUM): In scripts/download_video.py, the function check_yt_dlp executes pip install at runtime if the dependency is not found. This practice bypasses static analysis of dependencies and introduces a risk of Supply Chain Attack if the package registry is compromised.
[Command Execution] (MEDIUM): The skill relies heavily on subprocess.run to execute external commands. While it avoids shell=True, it passes user-provided URLs and output paths directly to the yt-dlp binary. A malicious user or an indirect prompt injection via a video URL could attempt to exploit flags in the underlying tool.
[Data Exposure] (LOW): The script allows the user to specify a custom output directory (-o). Without path validation or sandboxing, this could be used to write files to unintended locations on the host filesystem.
[Indirect Prompt Injection] (LOW): The skill fetches video metadata (title, uploader) from an external source (YouTube) and prints it to the console. An attacker could name a video with malicious instructions that might influence the LLM's subsequent actions when it parses the tool output.
Ingestion points: get_video_info fetches JSON metadata from YouTube.
Boundary markers: None. Metadata is printed directly to stdout.
Capability inventory: Subprocess execution of yt-dlp and pip, local file writing.
Sanitization: None implemented for the metadata display.

youtube-downloader