Source: Next-generation Constitutional Classifiers (Anthropic)
Anthropic (Jan 2026): an exchange classifier plus a cascade of activation probes cuts jailbreak-defence overhead to about 1% compute and benign refusals to 0.05%.
Contents
- What: research post (Constitutional Classifiers++)
- Publisher: Anthropic
- Date: 2026-01-09
- Original: anthropic.com/research/next-generation-constitutional-classifiers
Summary
Successor to the 2025 Constitutional Classifiers. Replaces separate input and output classifiers with an exchange classifier that judges an output in the context of its input, and arranges detection as a cascade: a cheap linear probe on the model’s internal activations screens all traffic, and only flagged exchanges go to a costlier probe-plus-classifier ensemble.
Key claims
- About 1% extra compute if applied to Claude Opus 4.0 traffic, versus 23.7% for the first generation.
- 0.05% refusal rate on harmless queries over a month of Claude Sonnet 4.5 traffic, 87% lower than the original system.
- Activation probes reuse computation the model already does and are harder to manipulate than output text.
- Red teaming: over 1,700 hours and 198,000 attempts found one high-risk vulnerability; no universal jailbreak found.
- The exchange classifier alone halved successful human red-team jailbreaks but cost 50% more compute, motivating the cascade.