gemma-tuner-multimodal
Pass
Audited by Gen Agent Trust Hub on Sep 14, 2026
Risk Level: SAFEREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [REMOTE_CODE_EXECUTION]: The skill clones a third-party repository (https://github.com/mattmireles/gemma-tuner-multimodal) and installs it as an editable package, which involves executing code from a non-trusted source. It also downloads model weights from Hugging Face.
- [COMMAND_EXECUTION]: Multiple shell commands are used to manage the virtual environment, install dependencies, and perform training operations via the custom gemma-macos-tuner CLI utility.
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to process external datasets (CSV files) for training, creating a vulnerability surface where data could influence agent behavior.
- Ingestion points: Reads training and validation samples from local CSV files specified in the configuration.
- Boundary markers: Absent; there are no instructions for the agent to distinguish between dataset content and operational commands.
- Capability inventory: The environment allows shell command execution, file system access, and network operations.
- Sanitization: No evidence of input validation or sanitization for the training data is provided in the documentation.
Audit Metadata