The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill instructs the agent to execute commands with elevated privileges using sudo to install system packages on Linux distributions (e.g., sudo apt update && sudo apt install -y yt-dlp).
[EXTERNAL_DOWNLOADS]: The skill facilitates the download and installation of external dependencies yt-dlp and openai-whisper. While openai-whisper is provided by a trusted organization (OpenAI), the skill still introduces external code into the environment via multiple package managers (pip, brew, apt).
[COMMAND_EXECUTION]: The skill constructs shell commands using dynamic output from yt-dlp, such as video titles, to create filenames and process data. The provided sanitization logic using tr is not exhaustive, leaving characters like backticks, semicolons, or dollar signs which could allow for command injection via malicious YouTube video titles.
[DYNAMIC_EXECUTION]: The skill uses python3 -c to execute inline Python code for processing and deduplicating subtitle files at runtime.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from YouTube video titles and transcripts, creating a surface for potential injection attacks.
Ingestion points: Untrusted data enters the agent context through the video title metadata and the content of the downloaded .vtt subtitle files.
Boundary markers: There are no explicit boundary markers or instructions to the model to ignore potential instructions embedded within the transcript text.
Capability inventory: The skill has access to shell execution (Bash), file writing, and local Python execution.
Sanitization: The skill attempts basic sanitization of video titles using tr, but this is insufficient to prevent all forms of shell-based injection.

youtube-transcript