The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Fetches source code from the official LLaVA GitHub repository and model weights from Hugging Face (liuhaotian/llava-v1.5-7b, openai/clip-vit-large-patch14-336). These are trusted, well-known locations for AI research and development.
[COMMAND_EXECUTION]: Provides instructions for standard development operations including repository cloning, pip installation, and running inference scripts. It also includes training commands using DeepSpeed for fine-tuning.
[INDIRECT_PROMPT_INJECTION]: As a multimodal assistant, the skill ingests external image data and text prompts, creating a surface for indirect injection. 1. Ingestion points: Processes local images using PIL and transformers in SKILL.md, and instruction datasets in references/training.md. 2. Boundary markers: Uses <image> and DEFAULT_IMAGE_TOKEN markers, but lacks explicit protection against adversarial instructions embedded in image pixel data. 3. Capability inventory: Includes capabilities for file system writes (model checkpoints) and network operations (downloading models). 4. Sanitization: No specific filtering or sanitization is implemented for the content of processed images or visual instructions.

llava