llava
Pass
Audited by Gen Agent Trust Hub on Oct 5, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: Fetches source code from the project's official GitHub repository and downloads pre-trained weights from Hugging Face model hubs.
- [COMMAND_EXECUTION]: Instructs users to install dependencies via pip and execute model training and inference scripts using standard CLI patterns.
- [INDIRECT_PROMPT_INJECTION]: Ingests image data for analysis, which serves as a potential entry point for adversarial instructions embedded in visual content.
- Ingestion points: Image files (image.jpg, document_image) processed by the vision-language pipeline.
- Boundary markers: Employs standard conversation templates to separate user and assistant roles.
- Capability inventory: Script execution for model inference and training, file system access for image/model loading, and network access for weight retrieval.
- Sanitization: Relies on the underlying model's inherent safety tuning and standard preprocessing.
Audit Metadata