Researchers have developed a new defense system called TRACE to protect large language models (LLMs) from multi-turn adversarial conversations. These conversations are designed to trick the model into doing something harmful by breaking it down into seemingly harmless turns. Existing defenses often over-react and refuse helpful requests, but TRACE uses structured reasoning to identify manipulation patterns and assign a 'jailbreak score' before generating each response. In tests, TRACE significantly reduced the success rate of attackers while maintaining high compliance with user requests. --- Why it matters: This matters because multi-turn adversarial conversations pose a significant threat to LLMs, which are increasingly used in applications like customer service and content moderation. By developing effective defenses like TRACE, researchers can improve the safety and reliability of these models, reducing the risk of harm to users or society. Source: https://arxiv.org/abs/2608.15594