LoD: New Jailbreak Detection for LVLMs Requires No Attack Samples

Researchers from Renmin University and Alibaba introduced LoD at EMNLP'26, a new framework for detecting jailbreak attacks in Large Vision-Language Models. By learning internal safety patterns from benign inputs rather than relying on known attack data, LoD identifies unseen attacks as anomalies. The method demonstrated superior performance across multiple LVLMs compared to existing baselines.

+3