Re-Triggering Safeguards within LLMs for Jailbreak Detection

A new research paper published on arXiv introduces a novel method for detecting jailbreaking prompts in Large Language Models (LLMs). Despite existing built-in safeguards, attackers can still craft prompts that bypass these defenses. The authors argue that such jailbreaking prompts are inherently fragile. To address this, they propose an embedding disruption technique designed to re-activate the LLM's internal safety mechanisms rather than replacing them with standalone solutions. The study includes a comprehensive analysis of disruption effects and presents an efficient search algorithm to identify optimal disruptions for detection. Extensive experiments demonstrate that this approach effectively defends against state-of-the-art jailbreak attacks in both white-box and black-box settings. Furthermore, the method remains robust even when facing adaptive attacks. This collaborative defense strategy enhances the security of AI systems by leveraging and reinforcing existing internal protections, offering a significant advancement in safeguarding LLMs against malicious manipulation and ensuring more reliable AI interactions.