Refusing Safe Prompts for Multi-modal Large Language Models