Anthropic says safeguards reduce extraction of Claude training data

Anthropic estimates that new safeguards cut reasoning tokens extracted from Claude by 10–100 times and says it banned tens of thousands of accounts.

The company announced the changes on October 12. Reuters reported that Anthropic made proxy networks harder to use for repeated requests and restricted access to Claude's reasoning.

Distillation trains a model on another system's responses to reproduce some of its capabilities. Anthropic characterises the campaigns it described as violations of its rules and attempts to copy protected Claude capabilities.

The company put the protection system's false-positive rate below 0.1%. These are its own measurements. The account does not provide an independent effectiveness assessment or a complete calculation methodology.

Anthropic named Chinese firm Z.ai, also known as Zhipu, among the alleged participants. It said the firm tried to use Claude responses while developing GLM-5.3. Reuters reported that Z.ai did not respond to a request for comment.

Anthropic, OpenAI and Google link the issue to competitors reproducing advanced capabilities more quickly and cheaply. Anthropic's announcement concerns specific safeguards and does not establish the legal status of every distillation practice.