AI and the World's Hard Problems · Theoretical computer science · Anthropic · 2026-10-09
Constitutional Classifiers
Large language models can still be jailbroken, meaning users bypass their safety guardrails to elicit harmful outputs. Anthropic's Safeguards Research Team proposed a defense based on Constitutional Classifiers, which are input and output classifiers. A prototype withstood thousands of hours of human red teaming for universal jailbreaks but had high overrefusal rates and compute overhead. An updated version showed similar robustness on synthetic evaluations, with only a 0.38% rise in refusal rates and moderate extra compute.