Anthropic accuses Alibaba of systematic model distillation theft

Anthropic, the developer of the Claude series of large language models, has publicly accused Chinese tech giant Alibaba of orchestrating a systematic campaign to extract proprietary AI capabilities. In a filing reviewed by CNBC and Bloomberg, Anthropic alleges that Alibaba engaged in 'brazen and illicit' model distillation—a technique where an attacker sends millions of queries to a target model through its API and uses the responses to train a competitive model. The accusation specifically points to Alibaba's Qwen series, suggesting that some of its advanced capabilities were derived from Claude models.

The move escalates a growing industry battle over IP theft in AI. Model distillation is technically difficult to prove, but Anthropic claims it has evidence of API usage patterns consistent with a large-scale extraction operation. This case could have major implications for the AI industry: if proven, it would demonstrate that even tightly controlled APIs are vulnerable to such attacks. It also intensifies geopolitical tensions, as U.S. AI firms increasingly view Chinese competitors as threats. Alibaba has not yet commented on the specific allegations, but the case highlights the challenges of protecting frontier model IP in a globalized market.