Evolutionary Methods Rather Than Prompts: Evolutionary Synthesis of Large Language Model Jailbreak Attacks

Researchers propose EvoSynth, an autonomous multi-agent framework that shifts the optimization space of large language model jailbreak attacks from prompts to executable code. The framework uses a multi-agent system to autonomously build, evolve and execute code-based attack algorithms, and includes a code-level self-correction loop that can iteratively rewrite attack algorithms based on target model feedback and failed attempts. Experiments show that EvoSynth achieves an attack success rate of 85.5% against highly robust models such as Claude-Sonnet-4.5, an average attack success rate of 95.9% on the evaluated targets, and significantly higher diversity of generated attacks than existing methods. The authors have released the framework to support subsequent research.