gemma-tuner-multimodal

Warn

Audited by Gen Agent Trust Hub on Apr 8, 2026

Risk Level: MEDIUMEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill instructs the user to clone and install a software package from a third-party GitHub repository (mattmireles/gemma-tuner-multimodal) and install dependencies from unverified requirements files.
  • [REMOTE_CODE_EXECUTION]: By guiding the user to install the repository via 'pip install -e .' and run the 'gemma-macos-tuner' CLI, the skill promotes the execution of unverified remote code on the local system.
  • [COMMAND_EXECUTION]: The instructions require the user to execute numerous shell commands, including environment variable exports for Hugging Face tokens and Google Cloud Service Account paths.
  • [PROMPT_INJECTION]: The skill possesses a surface for indirect prompt injection through the datasets it processes.
  • Ingestion points: Reads training data from local CSV files, Google Cloud Storage (GCS) buckets, and BigQuery tables.
  • Boundary markers: No specific delimiters or safety instructions are defined to separate training data content from the agent's logic.
  • Capability inventory: The system executes training loops, performs file system operations for model checkpoints, and maintains network connections for dataset streaming.
  • Sanitization: There is no evidence of sanitization or filtering of the input data to prevent the model or agent from potentially obeying instructions embedded within the training samples.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Apr 8, 2026, 12:55 PM