attention-machine
Warn
Audited by Gen Agent Trust Hub on Oct 15, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONPERSISTENCEEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill relies on numerous system-level tools including ffmpeg, ffprobe, yt-dlp, and Playwright. It automates the installation of these dependencies using system package managers such as Homebrew, winget, and npm.
- [PERSISTENCE]: During environment setup, the agent is instructed to modify the user's shell profile (specifically ~/.zprofile) to include Homebrew in the system PATH. This persistent system modification is performed automatically by the agent.
- [REMOTE_CODE_EXECUTION]: The setup process involves executing a remote shell script from a well-known service (Homebrew) and installing packages via npx from the vendor's repository.
- [EXTERNAL_DOWNLOADS]: The skill fetches resources from external services at runtime, including font files from Google Fonts and machine learning models from Hugging Face.
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted transcription data from the ElevenLabs API (ingestion point: snap-cuts.mjs). It lacks explicit boundary markers or sanitization for this content while maintaining high capabilities (subprocess execution, system configuration), creating an attack surface for indirect prompt injection.
Audit Metadata