Anthropic has unveiled that its newly implemented protective measures have substantially curbed the number of Claude inference tokens being extracted from its platform by distillation attackers, with the reduction rate ranging from 90% to 99%. Attackers frequently exploit networks of fake accounts to amass model responses in large quantities, thereby replicating the model's capabilities without authorization. Complete inference records hold greater value for training purposes than final answers alone. The new measures encompass restricting the transfer of inference records, deploying a two-stage classifier to detect and block extraction requests, and imposing rate limits or outright bans on accounts that persistently attempt such actions. The classifier boasts an extremely low false positive rate, below 0.1%, and has already resulted in the banning of tens of thousands of accounts associated with these attacks. Typically, the banning process is completed within two hours of an attack's initiation. Furthermore, starting from September 24, the company has reinstated charges for requests that are intercepted by its security defenses in specific categories. Some developers may find it necessary to update their integration methods. Additionally, there is a possibility that the classifier may inadvertently block legitimate requests. At present, the duration of the protective effectiveness remains uncertain, and the company has yet to encounter highly sophisticated and persistent bypass attempts. This underscores the need for continuous iteration and refinement of defensive measures.