Anthropic, both accuser and accused: the complaint against Alibaba lands amid a regulatory storm

Few stories capture the contradictions of the moment so well. On June 10, 2026, Anthropic sent a letter to the Senate Banking Committee, addressed to Tim Scott and Elizabeth Warren, accusing Alibaba and its AI lab of a "brazen" and "illicit" campaign —in the company's own terminology— to extract the capabilities of its models. The content was verified by CNBC and first reported by Bloomberg. The figures cited: 28.8 million exchanges between April 22 and June 5, with some 25,000 fraudulent accounts. All of it attributed, let us stress, by one of the parties.

The technique in question, distillation, is not illegitimate in itself: training a small model with the outputs of a larger one is a common practice in machine learning. The problem, according to Anthropic, arises when it is done at industrial scale, on third-party systems, without authorization and through fake accounts to evade controls. At that point it stops being an optimization and becomes, in its reading, a violation of terms of service and a threat to U.S. national security.

The case does not arise in isolation. In February 2026 the company had already flagged three "industrial-scale" campaigns attributed to DeepSeek, Moonshot, and MiniMax, and around April the White House Office of Science and Technology Policy issued a memorandum committing to help detect and coordinate responses. Anthropic maintains that Alibaba pressed ahead despite those warnings, which adds a dimension of deliberate defiance to its account. The narrative, in short, is one of escalation.

But the most revealing angle is the one that shows Anthropic in the opposite role. In the same weeks, the company disclosed that it had received an export control directive ordering it to suspend access to its latest models —cited as Fable 5 and Mythos 5— for any foreign national, inside or outside the U.S., including its own foreign employees, invoking "national security authorities" without further specifics. Its executives traveled to Washington to negotiate, and the company stated that "both sides are working quickly to resolve this." Accuser on one side, restricted on the other: AI leadership is starting to be paid for in regulatory friction as well.

For anyone building agentic systems on top of third-party models, there is a practical lesson that transcends the geopolitical tussle. The distillation Anthropic describes does not require access to the model's weights or architecture: it is enough to query it massively and learn from its responses. That turns the access layer —authentication, abuse detection, usage limits— into a security frontier as critical as training itself. The episode, far from being merely a clash between powers, is a reminder that in the agentic era value lies in behaviors that are, at once, extremely valuable and surprisingly easy to observe.