llava
Pass
Audited by Gen Agent Trust Hub on Sep 16, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: Fetches the LLaVA source code from the official GitHub repository (haotian-liu/LLaVA) and downloads pre-trained weights from the Hugging Face Hub (liuhaotian).
- [COMMAND_EXECUTION]: Utilizes shell commands to clone the repository, install Python dependencies in editable mode, and launch CLI or Gradio-based interfaces.
- [REMOTE_CODE_EXECUTION]: Executes shell scripts and training scripts included in the cloned repository to facilitate model pre-training and fine-tuning workflows.
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted image files and text queries through the vision-language model, creating a potential vector for indirect prompt injection.
- Ingestion points: External data enters the context via
Image.openand conversation prompt construction inSKILL.md. - Boundary markers: Employs standard
llava_v1conversation templates to delineate between human and assistant roles. - Capability inventory: The skill possesses capabilities to execute shell scripts and interact with the file system during training and inference.
- Sanitization: No explicit sanitization or filtering of multimodal input is performed in the provided code examples.
Audit Metadata