The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill makes network requests to YouTube's caption endpoints and uses yt-dlp to download media and metadata from platforms including Instagram, TikTok, X, and Vimeo. While these are established platforms, the broad nature of the download capability presents an inherent risk surface.
[COMMAND_EXECUTION]: The scripts fetch_transcript.py and fetch_media.py frequently invoke external binaries such as yt-dlp, ffmpeg, ffprobe, and various Whisper transcription engines using subprocess.run. Although the commands are constructed as lists (avoiding shell interpolation) and utilize the -- separator to prevent flag injection from user-supplied URLs, the execution of these binaries with external inputs is a high-capability behavior.
[OBFUSCATION]: The browser script scripts/browser_media_links.js contains logic to decode metadata from Instagram and Facebook URLs using atob (Base64 decoding). This is used to parse the 'efg' parameter for bitrate and version tags. While functional for its stated purpose, the dynamic decoding of external URL parameters is a technique that can be used to hide malicious payloads.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a significant attack surface for indirect prompt injection as it ingests untrusted video titles, descriptions, and transcripts.
Ingestion points: Transcripts and metadata are fetched from external video platforms and processed in fetch_transcript.py and fetch_media.py.
Boundary markers: The skill attempts to mitigate injection by wrapping external content in Markdown blockquotes (> ) and using explicit headers.
Capability inventory: The agent has access to subprocess execution (binary calls) and file writing capabilities.
Sanitization: The skill employs sanitize_inline_text, yaml_quote, and Markdown table cell escaping to neutralize control characters in the output.
[METADATA_POISONING]: The skill's own documentation (specs/trust-hub-audit.md) acknowledges a previous critical security failure related to the presence of a blacklisted domain (codeguilds.dev). Automated scans confirm the domain still appears in the project's historical planning documents (specs/youtube-fetcher-v1-1-release-and-distribution.html), posing a risk of user or agent redirection to a malicious site.

youtube-fetcher