Jailbreaking Multimodal Large Language Models using Multi-Clip Video
Researchers have developed a new dataset, MCV SafetyBench, to test the vulnerability of multimodal large language models (MLLMs) to malicious inputs. The dataset, comprising 2,920 videos, reveals that MLLMs are more susceptible to harmful content when presented with diverse, dynamic video inputs compared to static images. This research also highlights that the success rate of jailbreaking attacks increases with the number of video clips used, leading to the proposal of a defense strategy leveraging image modality robustness. AI
IMPACT Highlights potential security risks in video-processing AI and suggests new defense strategies.