The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses yt-dlp to fetch subtitles via shell commands. It interpolates a user-provided <URL> directly into the command string. If the agent does not properly escape or sanitize this URL, it could be exploited for command injection (e.g., using backticks or semicolons in the URL).
[EXTERNAL_DOWNLOADS]: The skill performs network requests to download subtitle files (.vtt) from external platforms.
[PROMPT_INJECTION]: The skill processes untrusted transcript data from external videos, creating a surface for indirect prompt injection where malicious instructions hidden in subtitles could influence the agent's behavior.
Ingestion points: Video URLs provided by the user and the content of downloaded .vtt files.
Boundary markers: None identified. The skill extracts and presents the text to the agent without clear isolation markers or "ignore instructions" warnings.
Capability inventory: Access to Bash and Read tools allows the agent to execute further commands or read local files if influenced by malicious content.
Sanitization: The vtt_to_paragraph.py script removes formatting tags and bracketed cues, but it does not perform dedicated security sanitization or validation of the text content for potential instructions.

video-transcript