The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill fetches video transcripts from external sources (YouTube and other sites) and prints them to the agent's context. This presents a surface for indirect prompt injection if a video author includes malicious instructions in the subtitles.
Ingestion points: scripts/vtd.js fetches transcripts via the youtube-transcript-plus library and yt-dlp subtitle downloads.
Boundary markers: No boundary markers or "ignore instructions" wrappers are used when printing the transcript content to stdout.
Capability inventory: The skill can execute shell commands (yt-dlp, ffmpeg) and write files to the local file system (outputDir).
Sanitization: The script performs basic cleaning of HTML tags and bracketed cues (e.g., [Music]), but does not sanitize the content for potential AI instructions.
[COMMAND_EXECUTION]: The vtd.js script allows forwarding arbitrary arguments to the yt-dlp binary using the -- separator. While it uses spawn with an argument array to prevent shell injection, yt-dlp itself contains flags (such as --exec) that allow for arbitrary command execution. If an agent is influenced to use these flags, it could lead to unauthorized system access.
[EXTERNAL_DOWNLOADS]: The skill is designed to download content from remote URLs provided by the user or agent, utilizing the youtube-transcript-plus package and the yt-dlp command-line utility. These operations are essential to the skill's primary purpose.

video-transcript-downloader