From the source
Anthropic says new anti-distillation methods cut the data harvested by distillers by around 90 to 99%, as it hardens Claude's reasoning against covert extraction campaigns.
- The company describes distillation attacks as industrial-scale efforts, often run through networks of fraudulent accounts, to copy a frontier model's capabilities into another model without authorization.
- Anthropic says it has turned away meaningful revenue and trillions of tokens of paying traffic from distillers, and that some changes forced legitimate users to update their integrations.
- Its defenses focus almost entirely on Claude's reasoning transcripts, the most valuable training data for distillers; summarizing reasoning before showing it to users dropped a distilled model's SWE-bench Verified score from 80.4% to 71.8%.
- After that change, attackers moved to active extraction, prompting Claude to reveal full reasoning, and Anthropic says its first safeguards were circumvented by proxy networks controlling hundreds of thousands of fraudulent accounts.
- Anthropic cites Zhipu, which it says distilled cyber capabilities during GLM-5.3's development and switched away from Claude Fable 5 after its cyber safeguards degraded the attacks, and a case in which Kimi requests were secretly forwarded to Claude.
