Research · curated 16 Sep 2026
MazeBreaker: Multi-Agent Reinforcement Learning for Dynamic Jailbreaking of LLM Security Defenses | Proceedings of the 2026 IEEE/ACM 48th International Conference on Software Engineering
First reported acm.org
Coverage timeline
Single-source research — first reported, latest, and curated coincide.
Why it matters
MazeBreaker demonstrates an adaptive, black-box RL-driven jailbreak that defeats strong commercial safety alignment, giving defenders a concrete new class of automated attack to test their LLM guardrails against.
MazeBreaker is a research paper (2026 IEEE/ACM ICSE) introducing a multi-agent reinforcement-learning jailbreak method that dynamically learns from feedback on target LLMs to bypass their safety alignment without a reference oracle. The authors evaluate it against six state-of-the-art jailbreak methods across 13 open-source and commercial models, reporting strong attack effectiveness against safety-aligned commercial models including GPT-3.5-turbo, GPT-4o-mini, GLM-4-air, and Claude-3.5-sonnet, with code and data released on GitHub.