llava
Pass
Audited by Gen Agent Trust Hub on Oct 2, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: Fetches the LLaVA source code from the project's official GitHub repository and downloads pretrained model weights from Hugging Face. These are legitimate resources for the LLaVA research project.
- [COMMAND_EXECUTION]: Provides instructions for installing the package via pip, running local inference via CLI, and launching a Gradio web interface for interaction.
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted image data and user text through a vision-language pipeline.
- Ingestion points: Image files (e.g.,
image.jpg) and text queries processed inSKILL.md. - Boundary markers: Employs
conv_templates(such asllava_v1) to maintain role-based separation between user input and model instructions. - Capability inventory: Includes local inference, command-line operations, and Gradio-based web serving.
- Sanitization: Relies on established LLaVA and Hugging Face image/token processing pipelines to handle multimodal input.
Audit Metadata