audiocraft-audio-generation
Fail
Audited by Gen Agent Trust Hub on Jun 18, 2026
Risk Level: HIGHCOMMAND_EXECUTIONDATA_EXFILTRATIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill provides instructions to execute commands with administrative privileges using 'sudo' to install system dependencies such as FFmpeg in the troubleshooting guide.
- [DATA_EXFILTRATION]: The Gradio demo implementation in 'SKILL.md' and 'advanced-usage.md' uses the 'share=True' parameter. This action creates a public URL and an unauthenticated tunnel to the local machine, exposing the service and potential internal environment to the internet.
- [EXTERNAL_DOWNLOADS]: The skill includes instructions to install packages and clone repositories from GitHub repositories associated with Meta Research (facebookresearch) and download models from HuggingFace.
- [COMMAND_EXECUTION]: Documentation in 'advanced-usage.md' includes instructions for using the 'dora' and 'git' command-line tools for experiment management and code retrieval.
- [PROMPT_INJECTION]: The skill processes untrusted user prompts for audio generation across its API server, Gradio interface, and LangChain tool components. Ingestion points: 'prompt' field in the FastAPI 'GenerateRequest' and the LangChain 'MusicGeneratorTool' description input. Boundary markers: None identified. Capability inventory: Subprocess execution ('pip', 'git', 'dora'), file system writes ('torchaudio.save'), and network operations (FastAPI server, Gradio tunnel). Sanitization: No input validation or prompt sanitization is documented before processing by generative models.
Recommendations
- AI detected serious security threats
Audit Metadata