Constitutional Classifiers++: Efficient Production-Grade Defenses against Universal Jailbreaks
This work addresses the critical need for practical and efficient safeguards against jailbreaks in large language models, offering a production-grade solution with significant improvements over previous defenses.
The paper tackled the problem of defending large language models against universal jailbreaks by introducing enhanced Constitutional Classifiers, which achieved a 40x reduction in computational costs and maintained a 0.05% refusal rate on production traffic while providing strong protection.
We introduce enhanced Constitutional Classifiers that deliver production-grade jailbreak robustness with dramatically reduced computational costs and refusal rates compared to previous-generation defenses. Our system combines several key insights. First, we develop exchange classifiers that evaluate model responses in their full conversational context, which addresses vulnerabilities in last-generation systems that examine outputs in isolation. Second, we implement a two-stage classifier cascade where lightweight classifiers screen all traffic and escalate only suspicious exchanges to more expensive classifiers. Third, we train efficient linear probe classifiers and ensemble them with external classifiers to simultaneously improve robustness and reduce computational costs. Together, these techniques yield a production-grade system achieving a 40x computational cost reduction compared to our baseline exchange classifier, while maintaining a 0.05% refusal rate on production traffic. Through extensive red-teaming comprising over 1,700 hours, we demonstrate strong protection against universal jailbreaks -- no attack on this system successfully elicited responses to all eight target queries comparable in detail to an undefended model. Our work establishes Constitutional Classifiers as practical and efficient safeguards for large language models.