Anthropic Accuses Alibaba of Illicitly Extracting Claude's Capabilities

Anthropic has accused Alibaba of illicitly extracting the capabilities of its Claude models, according to Reuters. The technique at issue is model distillation, in which one party queries a target model through its application programming interface (API) and trains a smaller student model on the responses, transferring behavior without access to the original weights.

What is verified here is narrow. Anthropic has made the allegation publicly. In the reporting available, the company has not published the query logs or statistical fingerprints that would let an outside party reproduce the claim, and Alibaba's account of how it trained its models has not been presented alongside it. Distillation from a competitor's API is difficult to prove and easy to deny, because a student model trained on broadly similar data can arrive at similar outputs without ever copying them. Readers should treat the accusation as an assertion from an interested party until independent evidence appears.

The timing connects two developments. A paper posted the same day, What Does It Mean to Break a Distillation Defense?, examines the output-perturbation defenses that API providers add to frustrate exactly this kind of copying. It argues that several proposed defenses can be circumvented, leaving the student model's accuracy largely intact. If distillation is cheap and the countermeasures are weak, then the value a lab captures from a frontier model can be extracted as soon as it is exposed through an API.

That economic reality is already shaping policy. Anthropic itself is operating under a US export-control directive that suspended foreign access to two of its models, and Washington has been extending controls from chips toward model access itself. An accusation that a leading Chinese firm distilled an American model strengthens the argument for restricting who can call frontier APIs at all.