Anthropic upgrades Claude's anti-distillation defenses, reducing model inference data leakage by 90% to 99%

PANews October 13 news: According to an official announcement by Anthropic, the company, through measures such as restricting access to reasoning data, deploying a two-stage detection system, and banning violating accounts, has reduced the amount of reasoning data illegally distilled and extracted from Claude by 90% to 99%, with a false positive rate below 0.1%. Tens of thousands of related accounts have currently been banned. Anthropic also alleged that Zhipu AI distilled the cybersecurity capabilities of leading U.S. models during the development of GLM-5.3, and called on the industry to jointly strengthen protections.