Large Language Models (LLMs)

We propose a multi-agent, multi-turn jailbreak strategy that systematically bypasses LLM safety mechanisms by decomposing harmful queries into seemingly benign sub-tasks.