Evolving Jailbreaks: Automated Multi-Objective Evolutionary Attacks Generate Long-Tail Safety Bypasses Missed by Manual Red-Teaming

arXiv·medium signalXBlueskyLinkedInCopy link

New automated approach uses multi-objective evolutionary search to generate diverse, long-tail jailbreaks that simultaneously maximize attack success rate and semantic coverage while evading RLHF-trained safety defenses. The method systematically discovers alignment gaps that manual red-teaming misses by exploring the full attack distribution rather than sampling obvious vectors. In agent contexts where a single safety violation chains into tool calls and environment modifications, undiscovered jailbreaks carry compounded downstream risk.