Anthropic has enhanced its defense against distillation attacks and banned tens of thousands of related accounts.

Beating AI Insight News Brief: Anthropic released a statement saying preliminary evidence shows new security measures have reduced the number of Claude inference tokens extracted by distillation attackers from its platform by approximately 90% to 99%. The company noted attackers typically use networks of fake accounts to bulk-collect model responses, unauthorizedly replicating its capabilities, adding that full inference records are more valuable for training than final answers. The new measures include restricting the transfer of inference records between conversations, models, and accounts, deploying a two-stage classifier to identify and block extraction requests, and rate-limiting or banning accounts that repeatedly attempt such attacks. Anthropic added the classifier has a false positive rate of less than 0.1%, and it has banned tens of thousands of distillation-related accounts, often completing bans within two hours of the start of relevant attacks. Additionally, since September 24, the company has resumed charging for requests blocked by security protections in specific categories such as distillation. Anthropic acknowledged that some developers will need to update their integration methods, and the classifier may also incorrectly block legitimate requests. It remains unclear how long the protective effects will last, and the company has not yet encountered highly sophisticated, sustained circumvention attempts, so it will continue to iterate on its defense measures.