As an early piece of work, this benchmark systematically reveals that current Multimodal Large Language Models are susceptible to malicious attacks.
Safety of Multimodal Large Language Models on Images and Text