Research
A Self-Evolving Multi-Agent Framework Defense against LLM Jailbreak Attacks
Publication date not yet evaluated · Added · 2 reports arxiv.org
Page published
Coverage timeline
Why it matters
Adaptive, memory-based defenses that generalize across jailbreak families address the static nature of most LLM guardrails, giving defenders a way to reduce attack success rates without retraining models.
A research paper by Tongyan Hu and Bryan Hooi (National University of Singapore) proposes a self-evolving multi-agent test-time defense against LLM jailbreak attacks, built on a persistent cross-interaction rule memory that abstracts successful attacks into method-level rules capturing the structural attack wrapper. The mechanism operates entirely through external memory and prompting with no parameter updates, applies to open-weight and black-box models, and is evaluated across four black-box jailbreak families.