Multi-model multi-agent defense against LLM jailbreak attacks

The robustness of Large Language Models (LLMs) against adversarial attacks such as jailbreak techniques is critical for their safe deployment especially in the sensitive domains. These attacks exploit LLM flexibility to bypass safety protocols, generating harmful, misleading, or unauthorized outputs. Existing single-agent and static defense mechanisms often struggle to address the dynamic and evolving jailbreak vulnerabilities. This paper introduces a Multi-Model Multi-Agent Defense (MAMAD) system designed to defend jailbreak attacks on LLMs. The proposed architecture incorporates specialized agents—Input Analyzer, Intent Analyzer, and Output Analyzer—utilizing the unique strengths of distinct LLMs to collaboratively evaluating inputs and mitigating potential threats. The system demonstrates significant improvements in detection accuracy against diverse attack vectors, including adversarial suffixes, latent injections, and role-playing techniques. Evaluation results show that the MAMAD system achieves a detection accuracy of 94%, outperforming individual LLMs and single-model multi-agent systems. Furthermore, this work explores the trade-offs between computational and latency overhead and system efficiency between individual LLMs and multi-agent defense architectures. The findings demonstrates the capability of multi-agent systems to improve the safety and reliability of LLM applications, while maintaining minimal false positive rates and scalability.