The Agent Skills Directory

[COMMAND_EXECUTION]: The reference file references/youtube-transcript.md provides explicit instructions for the agent to use administrative privileges: sudo apt update && sudo apt install -y yt-dlp. Instructing an AI to execute commands with elevated permissions poses a significant security risk to the host environment.
[EXTERNAL_DOWNLOADS]: The references/scripts/requirements.txt file references versions of core libraries that do not exist in the official PyPI registry, such as torch==2.8.0 and pyannote-audio==4.0.4. Referencing non-standard or 'future' versions is a known indicator of supply chain attacks, such as dependency confusion, where an attacker might publish a malicious package with a higher version number to a public repository.
[COMMAND_EXECUTION]: The Python script references/scripts/whisper.py utilizes subprocess.run to invoke external system utilities like ffmpeg and ffprobe. While the implementation uses argument lists to minimize shell injection risks, the execution still relies on file paths and inputs that could be influenced by untrusted external data.
[EXTERNAL_DOWNLOADS]: The skill is configured to download transcription models from HuggingFace and fetch media content from YouTube. While these are well-known services, the skill's core functionality depends on processing these external, binary data streams.
[PROMPT_INJECTION]: The skill's workflow involving the transcription of untrusted audio and video content creates an entry point for indirect prompt injection. If the resulting text contains malicious instructions that are subsequently followed by the agent, it could lead to unauthorized actions.
Ingestion points: Audio and video files, along with YouTube URLs, are processed through yt-dlp and faster-whisper (documented in SKILL.md and reference files).
Boundary markers: None detected; the skill does not implement delimiters or warnings to isolate transcribed text from the agent's instructional context.
Capability inventory: The skill possesses capabilities for file system access, network operations for downloads, and the execution of system commands via subprocesses.
Sanitization: The script subtitle_convert.py performs basic cleaning of HTML entities and tags, but no semantic sanitization is applied to the transcript to filter out potential instructions.

transcript