The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill contains several PowerShell installation scripts (install-auto-editor.ps1, install-whisper-cpp.ps1, install-qwentts-cpp.ps1) that download binary executables and model files from external repositories on GitHub and HuggingFace. These downloads are verified using SHA256 hashes defined in the scripts.
[REMOTE_CODE_EXECUTION]: External binaries are installed into the local file system (typically %LOCALAPPDATA%) and subsequently executed. While verification is in place, these binaries are sourced from third-party community repositories (WyattBlue, ggml-org, ServeurpersoCom) rather than established software vendors.
[COMMAND_EXECUTION]: The skill relies on extensive execution of system commands (FFmpeg, FFprobe, git, cmake, ninja). Commands are generally invoked using argument lists to prevent shell injection, though complex build processes for components like qwentts.cpp involve secondary shell environments.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data, such as YouTube subtitles (via yt-dlp) and user-supplied scripts. The AI agent is instructed to read and translate this content, creating a surface for indirect prompt injection. The skill's ability to execute local binaries and modify the system PATH increases the potential impact of adversarial content.
[SAFE]: Downloaded model files from HuggingFace are verified against hardcoded SHA256 hashes, and the skill includes comprehensive path and tag escaping logic (escape_ass_text, Escape-FilterText) to mitigate injection risks in generated files.

youtube-video-editor