Open the Pandora's Box of LLMs: Jailbreaking LLMs through Representation Engineering
Jailbreaking techniques aim to probe the boundaries of safety in large language models (LLMs) by inducing them to generate toxic responses to malicious queries, a significant concern within the LLM community. While existing jailbreaking methods primarily rely on prompt engineering, altering inputs to evade LLM safety mechanisms, they suffer from low attack success rates and significant time overheads, rendering them inflexible. To overcome these limitations, we propose a novel jailbreaking approach, named Jailbreaking LLMs through Representation Engineering (JRE). Our method requires only a small number of query pairs to extract ``safety patterns'' that can be used to circumvent the target model's defenses, achieving unprecedented jailbreaking performance. Building upon these findings, we also introduce a novel defense framework inspired by JRE principles, which demonstrates notable effectiveness. Extensive experimentation confirms the superior performance of the JRE attacks and the robustness of the JRE defense framework. We hope this study contributes to advancing the understanding of model safety issues through the lens of representation engineering.
PDF Abstract
