Anthropic got hit by export rules nobody understands

Governing AI through opaque, ad hoc interventions is unsustainable, experts warn.

Governing AI through opaque, ad hoc interventions is unsustainable, experts warn.

Anthropic has spent much of this week fighting to get its newest AI models back online after the Trump administration abruptly ordered the company to cut access for all foreign nationals, including users inside the US and its own 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 order, but in a statement on its website, Anthropic said the government cited “national security authorities” to justify “an export control directive” on the models. (Anthropic also claimed that the government’s concerns about a “jailbreak” potentially used by groups linked to China to access its models did not allow users to circumvent all of the company’s safeguards.)

But why did the administration use export control rules to address this? Experts say the episode appears to be unprecedented, exposing an uncertain and unstable stage in AI governance. And what, exactly, is Anthropic supposed to be exporting? (The company did not respond to The Verge’s request for comment.)

Export controls have traditionally applied to things that can be shipped across borders: weapons, hardware, tools, that kind of thing. Over time, the framework has expanded to cover less tangible goods, such as software, source code, technical data, and even 3D-printed gun files. These are still discrete things that can be copied, downloaded, published, or otherwise handed over and taken, not simply used through a remote service like a chatbot. In the context of AI, President Joe Biden moved to control AI model weights — the core data that makes a model work that can be copied and run elsewhere — in this manner; this idea was swiftly abandoned by the Trump administration in the second term.

The Anthropic order does not fit neatly into this framework. There is no obvious transfer taking place: Mythos and Fable remain hosted on Anthropic’s servers, and users do not receive source code, model weights, or a copy of the model themselves, instead getting the chatbot’s responses to their queries. The export could be some specific information produced by the models, but it’s not clear why that would require disabling access to the entire system rather than just restricting part of it. It could also be access itself — though remote access to cloud services is a known gap in current export control regimes, one that Congress is already trying to close through legislation now moving through the Senate.

Hanna Dohmen, a senior research analyst at Georgetown University’s Center for Security and Emerging Technology, told The Verge it is “an open question” as to whether the order strains existing rules without seeing the precise language behind it. “In any case, this regulation is quite notable because, to my knowledge, this is the first time US export controls have been used to control access to an AI model in this way.”

“To say that this is an unsettled area of export control rule-making would be an understatement,” said Andrew Reddie, a professor at UC Berkeley’s Goldman School of Public Policy. He said that export control rules and other regimes like arms regulations give the government “wide latitude” to restrict access to certain goods. But “the equivocation by successive administrations regarding what the responsibilities of model developers are” has made it hard for firms to understand what is expected of them, he said.

That leaves the industry in a bind. If Anthropic was targeted because Mythos and Fable are uniquely capable, the order raises obvious questions for the next generation of models from OpenAI, Google, Meta, xAI, and any other frontier lab. If they were targeted because of specific safeguard issues, the government needs to outline what protection it considers sufficient. And if Anthropic was singled out because of its testy relationship with the Trump administration, the order becomes even harder to make sense of.

Either way, experts say this is not a sustainable way to manage frontier AI, especially if the US wants to maintain its lead globally. The incident has already added fuel to arguments that governments and companies outside the US should be wary about relying on American firms for access to strategically important systems.

Reddie had similar concerns. “In some ways, I think this episode makes clear the unsustainability of the existing governance regime,” he said. That is especially true if the government was more concerned about whether users could jailbreak models and bypass their safeguards. “If creating models that are impossible to jailbreak becomes the de facto standard for the United States, then it will have no AI models.”

All of this points to the same problem: The Trump administration wants it both ways on AI. It has repeatedly said it wants to take a hands-off approach and champion American technology, yet forced a domestic champion to unceremoniously yank its frontier models through an order it has still not publicly explained. If Washington wants to control who can access powerful AI systems, it needs to say how, and give companies an actual chance of complying before launch. Ad hoc interventions seemingly delivered on a whim are not sustainable in the long run — and they are a good way to make sure the US falls behind in the AI race.

• Robert Hart