Researchers have invested considerable effort into ensuring that large language models (LLMs) align with human values, using various training techniques, such as instruction tuning and Reinforcement Learning from Human or AI Feedback (RLHF/RLAIF), to guard against text unsafety. However, these defenses remain incredibly vulnerable to some jailbreak attacks, which can cause the model to become overly defensive to sensitive topics or still generate harmful content, leaving the model performance particularly fragile. Therefore, to comprehensively study text safety and output robustness, we propose a latent jailbreak prompt dataset, each involving malicious instruction embedding. Specifically, we instruct the model to complete a regular task, such as translation, where the text to be translated contains malicious instructions. To further analyze the safety and robustness, we design a hierarchical annotation framework. We present a systematic analysis of the safety and robustness of LLMs concerning the position of explicit normal instructions, word replacement (verbs in explicit normal instructions, target groups in malicious instructions, cue words in malicious instructions), and instruction replacement (different explicit normal instructions). Our results show that current LLMs not only have a preference for certain instruction verbs, but also exhibit different jailbreak rates for different instruction verbs in explicit normal instructions. In other words, the probability of generating unsafe content by the model will be reinforced to varying degrees depending on the instruction verb in explicit normal instructions. Code and data are available at https://github.com/qiuhuachuan/latent-jailbreak.

翻译：暂无翻译