Anthropic has accused Alibaba Group of using thousands of fake accounts to illicitly extract capabilities from its Claude AI models.

This allegation highlights the intensifying competition between the U.S. and China to dominate artificial intelligence development. The ability to replicate advanced models without original research could shift the global balance of technological power and create significant security vulnerabilities.

In a letter sent to U.S. senators in early June 2026, Anthropic said that the Chinese technology conglomerate used 25,000 [1] fake accounts to bypass safeguards. These accounts were used to generate 28.8 million [2] synthetic conversations, which Anthropic said were used to distill and steal the proprietary logic of the Claude models.

According to the company, the activity took place on Alibaba's cloud-based AI platforms [3]. By generating these millions of interactions, the developer said Alibaba could effectively map the model's behavior to accelerate the creation of its own competing AI systems [4].

Anthropic said that such illicit extraction poses a competitive risk and a national security threat [4]. The company said that this process allows a foreign entity to benefit from the massive investments in research and compute power required to build the original Claude models [5].

These disclosures come amid heightened tensions regarding the global AI race. The issue gained further visibility around June 10, 2026, following remarks from Donald Trump regarding the competition with China [6].

Alibaba has not provided a public response to the specific figures cited in the Senate letter. The U.S. Senate has not yet announced formal hearings regarding the claims [3].

“Anthropic alleges that Alibaba illicitly accessed and extracted capabilities from its Claude AI models”

The dispute underscores a growing trend of 'model distillation,' where a smaller or competing AI is trained on the outputs of a more advanced one. If Anthropic's claims are verified, it suggests that traditional API safeguards are insufficient against state-level or conglomerate-scale coordinated attacks. This will likely lead to more aggressive 'walled garden' approaches to AI deployment and increased U.S. government scrutiny of how AI companies monitor foreign access to their frontier models.