The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/get_transcript.py executes the yt-dlp binary using subprocess.run(). This is implemented safely by passing arguments as a list rather than a shell string, which prevents command injection from malicious URL inputs.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection. It retrieves transcripts from YouTube, which are authored by third parties. If a transcript contains hidden instructions or malicious commands designed to manipulate the agent, the agent might follow them during summarization or analysis.
Ingestion points: scripts/get_transcript.py fetches external subtitle content via yt-dlp and prints it to stdout for agent consumption.
Boundary markers: None identified. The transcript text is provided to the agent without explicit delimiters or warnings to ignore embedded instructions.
Capability inventory: subprocess.run is used locally to run yt-dlp; the risk extends to other tools the agent may have access to when processing the transcript.
Sanitization: The clean_vtt function removes WebVTT metadata and timestamps but does not perform any natural language sanitization to detect or neutralize malicious instructions.

youtube-watcher