Jailbreak Attack
Jailbreak attacks exploit vulnerabilities in large language models (LLMs) and other AI systems, aiming to bypass safety mechanisms and elicit harmful or unintended outputs. Current research focuses on developing novel attack methods, such as those leveraging resource exhaustion, implicit references, or continuous optimization via image inputs, and evaluating their effectiveness against various model architectures (including LLMs, vision-language models, and multimodal models). Understanding and mitigating these attacks is crucial for ensuring the safe and responsible deployment of AI systems, impacting both the trustworthiness of AI and the development of robust defense strategies.
Papers
Unlocking Adversarial Suffix Optimization Without Affirmative Phrases: Efficient Black-box Jailbreaking via LLM as Optimizer
Weipeng Jiang, Zhenting Wang, Juan Zhai, Shiqing Ma, Zhengyu Zhao, Chao Shen
EEG-Defender: Defending against Jailbreak through Early Exit Generation of Large Language Models
Chongwen Zhao, Zhihao Dou, Kaizhu Huang
Hide Your Malicious Goal Into Benign Narratives: Jailbreak Large Language Models through Neural Carrier Articles
Zhilong Wang, Haizhou Wang, Nanqing Luo, Lan Zhang, Xiaoyan Sun, Yebo Cao, Peng Liu
Perception-guided Jailbreak against Text-to-Image Models
Yihao Huang, Le Liang, Tianlin Li, Xiaojun Jia, Run Wang, Weikai Miao, Geguang Pu, Yang Liu
Probing the Safety Response Boundary of Large Language Models via Unsafe Decoding Path Generation
Haoyu Wang, Bingzhe Wu, Yatao Bian, Yongzhe Chang, Xueqian Wang, Peilin Zhao