Anthropic reduced data harvested by distillers by 90 to 99%. Distillers copy frontier model capabilities without matching safeguards. Distilled models have been used for biology and cybersecurity harm. Zhipu distilled US models for GLM-5.3, which has weak safeguards. Distillers also target military and surveillance systems. Anthropic turned away trillions of tokens of paying traffic from distillers. Simple defenses like summarizing reasoning were bypassed. Distillers now use proxy networks with fraudulent accounts. New defenses include preserved thinking and two-stage classifiers. Classifiers have a false positive rate under 0.1%. Distillers now extract 10 to 100 times fewer reasoning tokens. Anthropic bans tens of thousands of distillation-linked accounts. Some legitimate users must update integrations due to preserved thinking.

• 90 to 99% reduction in harvested data
• 80.4% vs 71.8% SWE-bench Verified scores
• 10 to 100 times fewer reasoning tokens extracted
• <0.1% false positive rate