This research introduces Constitutional Classifiers - a novel approach to defend large language models against sophisticated jailbreak attacks that bypass safety guardrails.

• Trained on synthetic data generated using natural language rules (a "constitution")
• Successfully defended against jailbreaks across 3,000+ hours of red teaming
• Demonstrated effectiveness as a protective layer against multi-step harmful processes
• Provides scalable security that works across model variants

For security teams, this research presents a practical framework to enhance AI safety measures using language-based principles that can adapt to evolving threats, rather than static rules that attackers can learn to circumvent.

Constitutional Classifiers: Defending against Universal Jailbreaks across Thousands of Hours of Red Teaming