The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The download-video.sh script utilizes the yt-dlp --cookies-from-browser chrome flag. This command allows the skill to access and read the user's local browser cookies (sensitive credentials) to authenticate with video platforms like TikTok when standard access is blocked.
[EXTERNAL_DOWNLOADS]: The skill is designed to download media from arbitrary user-provided URLs. Furthermore, the fallback script download-fallback.sh interacts with a third-party service (greenvideo.cc) to extract video links, which are then downloaded using curl.
[COMMAND_EXECUTION]: The skill relies on executing bash scripts (download-video.sh, download-fallback.sh) that take user-controlled URLs as arguments. While the scripts use quoting, the handling of these URLs involves passing them to CLI tools like yt-dlp, curl, and agent-browser, creating a potential attack surface for command injection if inputs are not strictly validated.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external websites (videos and their associated metadata). The instructions explicitly suggest subsequent analysis of these downloaded files using tools like renoise analyze, which could lead to exploitation if the analysis tools are vulnerable to malicious content embedded in the media files.
Ingestion points: scripts/download-video.sh and scripts/download-fallback.sh via yt-dlp and curl.
Boundary markers: None identified; the skill processes raw external content.
Capability inventory: File system write (resources/references/), network operations (yt-dlp, curl, agent-browser), and browser interaction (agent-browser eval).
Sanitization: Basic regex checks are performed for platform detection, but the underlying media content is not sanitized.

video-download