A new front is opening in the geopolitical contest over artificial intelligence, shifting from hardware to the models themselves. According to a Bloomberg report, Anthropic has sent a letter alleging that Chinese tech giant Alibaba conducted a coordinated campaign between April and June 2026 using a technique known as “adversarial distillation.”
The accusation targets a method where one AI model is exploited to extract and replicate the capabilities of another, often in violation of terms of service. In adversarial distillation, an attacker systematically queries a target model—such as a large language model—with carefully crafted prompts, collects the outputs, and uses that data to train a competing model. This bypasses the need for original datasets or expensive training runs, effectively cloning proprietary intelligence.
Anthropic, a leading AI safety company, claims Alibaba employed this approach to accelerate its own model development. The letter, the full contents of which have not been published, frames the campaign as a deliberate and sustained effort to misappropriate Anthropic’s technology.
The backdrop is an escalating battle over AI dominance that has already seen restrictions on semiconductor exports and data center investments. Now, tensions are focusing on the models themselves—the core software assets—as state-linked actors seek shortcuts to close the capability gap. While adversarial distillation has been discussed as a theoretical risk, a formal allegation of this kind marks a significant escalation. If substantiated, it could trigger new rounds of regulatory scrutiny, export controls on model access, and heighten the already strained tech rivalry between the U.S. and China.