HoneyTrap: Outsmarting Jailbreak Attacks on Large Language Models

Researchers present HoneyTrap, a deceptive multi-agent defense framework that counters multi-turn LLM jailbreak attacks by intercepting, misdirecting, and prolonging adversarial interactions. Evaluated using the MTJ-Pro dataset and new metrics (Mislead Success Rate and Attack Resource Consumption), HoneyTrap significantly reduces attack success rates and increases attacker costs while maintaining helpfulness and user experience across major LLMs.