heartmula
Pass
Audited by Gen Agent Trust Hub on Oct 30, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill clones the repository 'HeartMuLa/heartlib' from GitHub and downloads model checkpoints from HuggingFace, which is a well-known service for hosting AI models.
- [COMMAND_EXECUTION]: Shell commands are used to create a virtual environment, install packages using 'uv', and execute the generation script.
- [DYNAMIC_EXECUTION]: The instructions include specific Python code blocks to patch the downloaded library, addressing RoPE cache issues and tensor size mismatches in the 'transformers' and 'HeartCodec' modules.
- [INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied lyrics and tags. Evidence chain: 1. Ingestion points: 'lyrics.txt' and 'tags.txt' are read by the generation script. 2. Boundary markers: No explicit delimiters or warnings are provided in the instructions. 3. Capability inventory: The script generates audio files (.mp3) based on the input. 4. Sanitization: No sanitization of input lyrics or tags is described.
Audit Metadata