Anthropic was affected by export regulations that no one understood

Anthropic has spent this week scrambling to bring its new AI models online in the wake of the Trump administration he suddenly ordered the company to limit access to all foreign nationals, including users within the US and its employees, forcing Anthropic to block access to Fable 5 and Mythos 5 for everyone.

The Trump administration has not publicly explained the legal basis for the plan, but in a statement on its website, Anthropic he said The government called on “national security authorities” to approve the “export control law” on samples. (Anthropic reported that government concerns about “jailbreak” which can be used by groups connected to China to find examples of it did not allow users to bypass all company security.)

But why did the authorities use foreign policy to deal with this? Experts say this event appears to be unprecedented, revealing an uncertain and unstable role in AI governance. And, exactly, should Anthropic be exported? (The company did not respond Seasiderequest for comments.)

Export controls have been used for items that can be shipped across borders: weapons, equipment, tools, things of that nature. Over time, this framework has expanded to cover more limited resources, such as software, source code, technology, and more. 3D printed gun files. These are still special things that can be copied, downloaded, printed, or given and taken away, not just used remotely like a chatbot. In the case of AI, President Joe Biden to be moved managing the weight of the AI ​​model – the big data that makes up the model work that can be copied and managed elsewhere – as follows; this idea was to be abandoned early and the Trump administration in the second phase.

The Anthropic system does not fit well with this framework. No automatic migration is taking place: Mythos and Fable remain hosted by Anthropic’s servers, and users do not receive source code, sample weights, or a copy of the sample itself, instead they get chatbot answers to their questions. Exporting may be a special feature created by the model, but it is not clear why it would want to restrict access to the entire system rather than restricting part of it. It can also be self-sufficient – although the remote availability of cloud services is a notable exception in the current movement, which Congress is already trying. nearby through legislation going through the Senate.

Hanna Dohmen, senior researcher at Georgetown University’s Center for Security and Emerging Technology, said Seaside it’s an “open question” whether the plan violates existing laws without seeing the precise language behind it. “In any case, this order is notable because, as far as I know, this is the first time that US authorities have been used to regulate access to an AI model in this way.”

“To say that this is an unsustainable role in foreign policymaking would be an understatement,” said Andrew Reddie, a professor at UC Berkeley’s Goldman School of Public Policy. He added that export controls and other controls such as arms give the government the “right” to restrict access to certain goods. But the “uniformity of successive jurisdictions regarding the role of model makers” has made it difficult for companies to understand what is expected of them, he said.

That leaves the industry in jail. If Anthropic is targeted because Mythos and Fable are uniquely capable, the system raises clear questions for the next generation of models from OpenAI, Google, Meta, xAI, and any other frontier lab. If they were targeted because of other security issues, the government must explain the security it deems to be adequate. And if Anthropic was chosen for that reason a definite alliance with the Trump administrationthe system is difficult to understand.

In any case, experts say this is not a sustainable way to manage the frontiers of AI, especially if the US wants to remain a global leader. The This issue has already added to the controversy that governments and companies outside the US need to be careful about relying on American companies for critical infrastructure.

Reddie had similar concerns. “In some ways, I think this case shows the instability of the current administration,” he said. This is especially true if the government is very concerned about whether users can hack prisons and bypass their security. “If creating models that are impossible for prisons to destroy becomes legal in the United States, then it will not have AI models.”

All of this points to the same problem: The Trump administration wants both ways on AI. It has he repeatedly said it wants to take an independent path and champion American manufacturing, but it has forced the domestic hero to simply freeze its peripheral models in a system that it has not publicly explained. If Washington wants to control who gets access to powerful AI systems, it needs to say how, and give companies a real chance to follow suit before they are implemented. A seemingly half-hearted intervention is unsustainable in the long run — and it’s a good way to ensure the US lags behind in the AI ​​race.