The government’s offensive trump to limit Anthropic PBC’s operations, embodied in a recent Commerce Department order, relies on unprecedented use of export control laws. It also raises legal questions about whether the US can determine Who has access to artificial intelligence systems.
By demanding that Anthropic obtain US authorization for foreign citizens to access its Fable 5 and Mythos 5 models, the Secretary of Commerce, Howard Lutnick took export control regulations into unprecedented territory: the use of advanced artificial intelligence models. The decision has raised concerns in the industry over the government’s growing scope for intervention in the name of national security.
Before the directive issued last week, the technology industry assumed based on previous official interpretations that export controls did not apply to the use of cloud programs, although they could restrict the transfer of their source code. Lutnick’s letter called that judgment into question. “That can no longer be stated,” he said. Kate Koren, deputy director of the Center for Strategic and International Studies (CSIS).
“Now that the Department of Commerce has done it, no company can rule out doing it again,” he added. Koren, who previously worked in that department. “As long as this measure is not challenged or revoked, any customer should assume that it could happen with any model and at any time.”
The order marked the government’s deepest intervention in the AI industry yet and led Anthropic to immediately suspend access to both models. It also marked a shift from the approach laid out just two weeks earlier in an executive order from Donald Trump, which relied on the voluntary participation of companies to evaluate the security of AI systems.
The letter sent by Lutnick to the company, reviewed by Bloomberg News, has opened a debate about whether the simple use of an AI model can legally be considered a technological transfer. That interpretation — key to supporting government-imposed restrictions — breaks with standard practice on access to software capabilities, according to export controls experts and international trade lawyers.
The Bureau of Industry and Security (BIS), responsible for overseeing US export control programs, has for years enforced rules requiring companies to obtain government authorization before sharing certain sensitive technologies, including with foreign employees, in a process known as “deemed exports.” However, extending that concept to the use of AI systems is uncharted territory.
“For the provisions cited in the letter, it matters a lot what the department means by ‘model,’” said Chris Chamberlain, a former Commerce Department counsel and partner in the national security group at the law firm Morrison Foerster. “The basic interpretation is that offering access to software through the cloud does not constitute an export or deemed export of that software.”
In his letter, addressed to Anthropic CEO Dario Amodei, Lutnick cited two legal export control powers that he claimed allow the U.S. to impose licensing requirements. However, these powers have been little used and could only apply to certain countries, rather than having a global reach.
The first rule allows the US to identify emerging technologies and impose temporary restrictions on them. The second empowers the government to quickly establish licensing requirements when there is a risk that controlled technologies will reach users linked to military intelligence in countries considered adversaries, such as China or Russia.
Although the government has not yet publicly explained the reasons for the decision, the company maintains that the order was issued after it was discovered that it was possible to “jailbreak”, that is, bypass the security barriers of Fable 5, a recently released version of Mythos that Anthropic had blocked for cybersecurity tasks.
“The US solved a problem,” Koren said, referring to the risk associated with jailbreaking. “But it opened a Pandora’s box by doing it this way.”
Since last week, the company’s technical staff and government officials have been holding talks to try to resolve national security concerns raised by Washington, although so far there have been few public signs of progress. Asked on Wednesday during the Group of Seven summit in France about the status of negotiations, trump He responded that they were going “well,” without offering more details.
At the G7 summit, the US offensive against Anthropic took center stage. French President Emmanuel Macron led talks on mechanisms to deploy advanced AI models through so-called trusted partners. Overall, the US move has “reignited calls around the world for technological sovereignty and the independence of the US ecosystem,” said Michelle Nie, a researcher at the Center for a New American Security.
For now, the Commerce Department’s measures require Anthropic to obtain a license to offer those models to anyone considered foreign. The company does not currently require verification of the citizenship of its users and, in addition, has foreign citizens among its own employees.
The move forces the rest of the AI industry to take seriously the authority Lutnick claims to have. Furthermore, if the provisions used against Anthropic are challenged, the BIS could use other powers to impose restrictions. This increases the risk that companies that clash with the US government’s demands will suffer interruptions in access to key services.
Aidan Gomez CEO of Canadian AI developer Cohere Inc., said Thursday that the episode demonstrates the need for companies to rely on a variety of suppliers.
“It’s exactly what we’ve been saying for a long time,” Gomez said in an interview during the VivaTech conference in France. “If you rely on a few large, centralized players, you risk losing access, which is why you need a diversified AI supply chain.”
