America’s new AI policy aims to restrict who has access to frontier models. It’s a tough challenge.
Since taking office, President Trump’s administration has slashed regulatory obstacles to promote the American AI stack in the name of innovation. It rolled back the Biden-era AI Diffusion Rule, which created a global licensing regime for access to advanced semiconductors; its AI Action Plan aimed to remove “onerous regulation”; and it moved to block state-level AI safety laws. The overall goal was to make US tech indispensable.
“The AI industry is a beautiful baby,” the President explained. “We have to grow that baby and let that baby thrive. And you don’t want rules to get in the way of a beautiful baby when it’s taking its first steps. You just want it to take its first steps, essentially, and let it start to run.”
But as AI models have grown potentially more dangerous, the administration has shifted gears, citing national security and real concerns about the model’s own security. The policy moves far beyond Biden-era tools to control how frontier models are accessed and by whom.
An executive order, dubbed Promoting Advanced Artificial Intelligence Innovation and Security, represented a first, baby step. Although it appears to validate concerns over AI safety, the administration abandoned early ideas of mandatory government checks in favor of a voluntary framework under which AI labs would give authorities access to frontier models 30 days before their release.
The real reversal came with export controls. Within ten days of the voluntary executive order, the Commerce Department ordered Anthropic to cut off all access to its latest Mythos 5 and Fable 5 models for non-US citizens. Amazon reportedly raised concerns about cybersecurity risks. Anthropic disagrees.
In any case, the order is unworkable. Anthropic could only comply by firing its own foreign employees. Software, by its nature of bytes and bits, easily crosses physical borders. In response to the administration move, Anthropic suspended access for everyone, including US citizens.
Allies are spooked. European leaders, already worried about their dependency on US tech, had just announced a tech sovereignty plan designed to limit their reliance on US cloud, chips, and AI models. The ban on Mythos and Fable access reinforced fears that the US government would wield a “kill switch” to limit access to cutting-edge AI.
The Anthropic case “further underlines Europe’s need for technological sovereignty,” European Commission spokesperson Thomas Regnier said. In a volatile European Parliament debate, even the most pro-tech voices called the ban on the Anthropic models a wake-up call.
“Europe cannot keep building its tech stack on access that can be switched off overnight by a foreign government,” said Aura Salla, a former Meta executive turned European parliamentarian. “We must take action to reserve our data and our market preliminarily for European tech to scale it and build our own frontier AI.”
From the US perspective, this crackdown has a logic. It reflects a growing recognition that semiconductor export controls have underdelivered. China has found other ways to access compute and train capable AI models. Allies who are central to the regime have proved difficult. US lawmakers recently introduced a bill to extend US export restrictions on Dutch and Japanese lithography machines, angering allied governments who are “wary of new export controls as a US legislative mandate rather than a negotiated arrangement,” according to Emily Benson and Lexi Linafelter of the Center for Strategic and International Studies.
The Anthropic case underlines an expansion of export controls to target software models, not just hardware. It is problematic. What threshold should count as a national security risk? A complete cutoff to foreigners is difficult to enforce in practice. Restrictions can also be circumvented, as China has shown by reportedly accessing Mythos.
Allied cooperation may be the only realistic solution. At the G7 summit in France, US Commerce Secretary Howard Lutnick pitched giving “trusted partners” privileged access to the latest frontier models.
This approach contains risks. Access to the American AI stack would then become a privilege that Washington can grant or remove at will, making allies scramble for a place on the “trusted” list.
Washington wants to make itself the gatekeeper of frontier AI models. But in restricting access, it encourages both US rivals and allies to pursue AI alternatives.
Marta Granados Hernández is a US Google Policy Fellow at the Center for European Policy Analysis (CEPA) and a master’s candidate at Georgetown University’s Master of Science in Foreign Service program, concentrating in Science, Technology, and International Affairs.
Bandwidth is CEPA’s online journal dedicated to advancing transatlantic cooperation on tech policy. All opinions expressed on Bandwidth are those of the author alone and may not represent those of the institutions they represent or the Center for European Policy Analysis. CEPA maintains a strict intellectual independence policy across all its projects and publications.
Tech 2030
A Roadmap for Europe-US Tech Cooperation
