llava
Pass
Audited by Gen Agent Trust Hub on Oct 29, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: Fetches the LLaVA source code from its official GitHub repository and retrieves pre-trained model weights from Hugging Face.\n
- Evidence:
git clone https://github.com/haotian-liu/LLaVAinSKILL.md.\n - Evidence: Downloads models from the
liuhaotianorganization on Hugging Face.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted image data and user text prompts which acts as a surface for indirect instructions.\n - Ingestion points: Image files (e.g.,
image.jpg) and text queries processed viaSKILL.md.\n - Boundary markers: Employs conversation templates and specific image tokens (
DEFAULT_IMAGE_TOKEN) to structure interactions.\n - Capability inventory: Focuses on model inference and training; no sensitive system commands or network exfiltration logic is present.\n
- Sanitization: Input content is processed directly by the model without an explicit pre-filtering or sanitization layer in the provided examples.
Audit Metadata