GuardrailAgent: An Agent Framework for Jailbreak Defense

The GuardrailAgent multi-agent jailbreak defense framework achieves an overall detection rate of 93.9% on more than ten representative jailbreak attacks, and continuously updates its knowledge base by automatically aggregating attack features, supporting both input-side and output-side detection. Its input-side detection rate is 93.1%, with an F1 of 95.8%, and it can adapt relatively quickly to unseen attacks.