The Rising Threat of Multi-Turn Jailbreaks in Large Language Models

Large Language Models (LLMs) are advancing at a breathtaking pace, yet their vulnerabilities are harder to dismiss. The latest threat? Multi-turn jailbreaks. While single-turn attacks have been extensively analyzed, the evolving complexity and adaptability of multi-turn interactions introduce new challenges.

Introducing PLAGUE: A Novel Framework

In the context of these sophisticated attacks, a new framework called PLAGUE emerges as a major shift. Drawing inspiration from lifelong-learning agents, PLAGUE dissects the lifecycle of a multi-turn attack into three distinct phases: Primer, Planner, and Finisher. This methodical approach allows for a more comprehensive exploration of how these attacks unfold.

The benchmark results speak for themselves. Evaluations show that agents using PLAGUE achieve state-of-the-art results, with attack success rates (ASR) increasing by over 30% compared to previous methods. Notably, on models like OpenAI's o3, PLAGUE achieves an impressive ASR of 81.4%. Even Claude's Opus 4.1, known for its strong defenses, succumbs with a 67.3% ASR.

Why This Matters

Why should we care about these numbers? As LLMs become increasingly integrated into our workflows and daily lives, their susceptibility to these nuanced attacks poses significant risks. The framework highlights important aspects of model vulnerability, emphasizing the need for reliable defenses.

Western coverage has largely overlooked this level of sophistication in multi-turn jailbreaks. The paper, published in Japanese, reveals insights that demand attention. The rising success rates of these attacks shouldn't be dismissed as mere technical footnotes. They represent a tangible threat that could undermine trust in AI systems we rely on.

A Call to Action

What does this mean for developers and policymakers? It's time to prioritize security in model development. As models grow more complex, so do the attacks against them. The responsibility lies not only with AI researchers but also with those who deploy these systems across industries.

Is the AI community prepared to address these vulnerabilities head-on? The data shows that a proactive approach is needed to stay ahead of potential attacks. As we strive for progress, let's ensure that security isn't left behind.