Adversarial Confusion Attack: Disrupting Multimodal Large Language Models