TROJail Framework Uses Reinforcement Learning to Optimize Multi-Turn Jailbreak Attacks on LLMs

A recent study presents TROJail, a novel technique designed to enhance multi-turn jailbreak attacks on large language models by utilizing reinforcement learning. This method treats jailbreak attacks as a multi-turn reinforcement learning challenge, aiming to optimize the harmfulness of the final output as the reward. To tackle the issue of sparse supervision from outcome rewards, TROJail incorporates two process rewards that assess intermediate prompts and factor them into advantage estimation. These rewards discourage excessively harmful prompts that may activate model refusal mechanisms while promoting semantic relevance toward desired responses. The research highlights vulnerabilities in popular large language models that are still at risk from multi-turn jailbreak attacks, jeopardizing their safe usage. Existing methods often depend on turn-level optimization, which is inadequate for developing long-term attack strategies. Identified as arXiv:2512.07761v3, this paper signifies a significant advancement in training automated multi-turn attackers to explore model safety weaknesses, bridging the divide between turn-level optimization and extensive attack strategy formulation.