According to CoinMeta, Anthropic stated that they have strengthened defenses against distillation attacks and banned tens of thousands of related accounts. Preliminary evidence shows that the new protective measures have reduced by about 90% to 99% the amount of token that distillation attackers can extract from their platform. The company mentioned that attackers typically use a network of fake accounts to bulk obtain model responses, unauthorizedly replicating their capabilities, with complete reasoning records being more valuable for training than the final answers. The new measures include restricting the transfer of reasoning records between conversations, models, and accounts, deploying a two-stage classifier to identify and intercept extraction requests, and imposing rate limits or bans on accounts that make repeated attempts. Anthropic claimed that the false positive rate of the classifier is less than 0.1%, and tens of thousands of accounts related to distillation have been banned, often within two hours of the start of such attacks. In addition, since September 24th, the company has resumed charging for requests that are intercepted by security measures for specific categories such as distillation.