Anthropic Says New Defenses Cut Distillers' Claude Reasoning Extraction 10 to 100 Times
Summary
Anthropic says its new anti-distillation defenses cut the reasoning tokens attackers extract from Claude by 10 to 100 times. The company's two-stage classifier, run on Claude Sonnet 5.5 and Opus 5.5, has a false positive rate under 0.1 percent. Anthropic also charges for requests blocked before Claude responds.
Key Points
- Anthropic says it bans tens of thousands of distillation accounts, often within two hours of a campaign starting.
- Anthropic reports that Zhipu distilled the cyber capabilities of leading US models during GLM-5.3's development.
- Preserved thinking blocks new API accounts, those created on or after Aug. 31, 2026, from editing Claude's prior context.