'Constitutional Classifiers' Technique Mitigates GenAI Jailbreaks

Anthropic researchers propose "Constitutional Classifiers," a synthetic-data trained input/output filtering approach to block LLM jailbreaks. In extensive human red-team testing, the classifiers reduced jailbreak success from 86% to 4.4% while minimally increasing refusals and adding moderate compute overhead, aiming to prevent malicious extraction of dangerous CBRN and other sensitive information.