The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/download.sh script executes shell commands such as mkdir and yt-dlp. It employs secure bash array syntax (ytdlp_cmd) to construct and execute commands, which effectively prevents shell injection from user-supplied arguments.
[EXTERNAL_DOWNLOADS]: The skill includes an automatic installation routine for the yt-dlp utility, utilizing trusted system package managers including Homebrew (brew) and Python's pip3. These are standard methods for managing dependencies in this context.
[INDIRECT_PROMPT_INJECTION]: The skill processes data from external YouTube servers, specifically video titles and metadata, to generate local file paths. This metadata is untrusted and could serve as an injection vector if an agent later reads and acts upon the downloaded file names or metadata.
Ingestion points: The url_or_id parameter and the resulting external metadata fetched by the yt-dlp process in scripts/download.sh.
Boundary markers: The skill does not implement specific boundary markers or instructions to the agent to ignore content within the processed external data.
Capability inventory: The skill possesses network access and file system write capabilities.
Sanitization: Employs regex-based validation in the normalize_url function to ensure that user-provided identifiers conform to expected YouTube ID or URL formats.
[DATA_EXFILTRATION]: The documentation suggests the use of the --cookies-from-browser flag to access age-restricted content. While this is a standard feature of the underlying tool, it requires granting the utility access to sensitive browser session data, which carries inherent risks of data exposure if misused.

youtube-downloader