Blab AI
U.S. Commerce Order Forces Anthropic to Suspend Foreign Access to New AI Models
A letter from the U.S. Department of Commerce on June 12 sent a clear warning to Anthropic PBC: the company must secure a U.S. license before any foreign national can use its Claude Fable 5 or Claude Mythos 5 models. The directive, signed by Commerce Secretary Howard Lutnick, expands export‑control rules to cover cloud‑based artificial‑intelligence models—an area that had previously been treated as exempt.
For months, industry observers had believed that export controls applied only to the transfer of software code, not to the act of running a model hosted in the cloud. Kate Koren, a deputy director at the Center for Strategic and International Studies, said the letter “puts an end to that assumption.” She added that, “until or unless this is challenged, any customer has to assume this could happen for any model at any time.”
The order followed a brief window in which Anthropic had rolled out Fable 5 and Mythos 5 earlier in June. The company immediately disabled public access to both models, citing the Commerce Department’s directive. The shutdown marked the most significant government intervention in the AI sector since the Trump administration’s executive order on AI safety issued in early May.
The letter cites two U.S. statutes that give the government authority to impose export‑control licenses. The first allows the U.S. to identify emerging technologies and impose interim restrictions. The second permits rapid licensing to prevent controlled technology from reaching military intelligence users in adversarial states such as China or Russia. The applicability of these statutes to AI usage, however, remains largely untested.
Bureau of Industry and Security (BIS) officials have long required companies to secure approval before exporting sensitive technology—a process known as “deemed exports.” Chris Chamberlain, a former Commerce adviser and partner at Morrison Foerster, noted that the baseline interpretation is that providing cloud access to software does not constitute an export or deemed export. The Anthropic order challenges that interpretation.
The directive also raises a legal question: does the mere use of an AI model count as a technology transfer? Export‑control analysts and trade lawyers say the order’s position runs counter to past practice regarding access to software capabilities.
Anthropic’s CEO, Dario Amodei, received the letter. The company believes the order was prompted by the discovery that the Fable 5 model could be “jailbroken,” allowing it to bypass guardrails that were intended to block cybersecurity tasks. Kate Koren described the situation as a “can of worms” opened by the government.
The issue gained international attention during the Group of Seven (G‑7) summit in Paris. French President Emmanuel Macron led discussions on deploying advanced AI models through “trusted partners.” Michelle Nie, a fellow at the Center for a New American Security, said the U.S. move “renewed worldwide calls for technological sovereignty and independence from the American ecosystem.”
For now, the Commerce Department’s order requires Anthropic to obtain a license before offering the models to any foreign person. Anthropic does not currently verify citizenship for its users, and it employs foreign nationals. The order signals that other AI companies may face similar scrutiny, and BIS could invoke additional powers if the Anthropic case is contested.
Aidan Gomez, CEO of Canadian AI developer Cohere, said the episode highlights the need for diversified AI supply chains. “If you rely on centralized large players, you are at risk of losing access,” he said at the VivaTech conference in Paris.
The directive has not yet been challenged in court, and the administration has not provided a detailed rationale. Anthropic’s technical and administrative teams are reportedly in talks with U.S. officials, but no resolution has been announced.
The situation underscores a broader trend of U.S. export‑control policy being applied to high‑tech software and AI. The policy shift could affect the availability of advanced models for researchers, businesses, and governments outside the United States, and it may prompt other countries to reassess their own AI export controls.
As the industry watches, the next steps will likely involve legal challenges, negotiations for licensing, and potential adjustments to export‑control frameworks to clarify the status of cloud‑based AI usage.
For months, industry observers had believed that export controls applied only to the transfer of software code, not to the act of running a model hosted in the cloud. Kate Koren, a deputy director at the Center for Strategic and International Studies, said the letter “puts an end to that assumption.” She added that, “until or unless this is challenged, any customer has to assume this could happen for any model at any time.”
The order followed a brief window in which Anthropic had rolled out Fable 5 and Mythos 5 earlier in June. The company immediately disabled public access to both models, citing the Commerce Department’s directive. The shutdown marked the most significant government intervention in the AI sector since the Trump administration’s executive order on AI safety issued in early May.
The letter cites two U.S. statutes that give the government authority to impose export‑control licenses. The first allows the U.S. to identify emerging technologies and impose interim restrictions. The second permits rapid licensing to prevent controlled technology from reaching military intelligence users in adversarial states such as China or Russia. The applicability of these statutes to AI usage, however, remains largely untested.
Bureau of Industry and Security (BIS) officials have long required companies to secure approval before exporting sensitive technology—a process known as “deemed exports.” Chris Chamberlain, a former Commerce adviser and partner at Morrison Foerster, noted that the baseline interpretation is that providing cloud access to software does not constitute an export or deemed export. The Anthropic order challenges that interpretation.
The directive also raises a legal question: does the mere use of an AI model count as a technology transfer? Export‑control analysts and trade lawyers say the order’s position runs counter to past practice regarding access to software capabilities.
Anthropic’s CEO, Dario Amodei, received the letter. The company believes the order was prompted by the discovery that the Fable 5 model could be “jailbroken,” allowing it to bypass guardrails that were intended to block cybersecurity tasks. Kate Koren described the situation as a “can of worms” opened by the government.
The issue gained international attention during the Group of Seven (G‑7) summit in Paris. French President Emmanuel Macron led discussions on deploying advanced AI models through “trusted partners.” Michelle Nie, a fellow at the Center for a New American Security, said the U.S. move “renewed worldwide calls for technological sovereignty and independence from the American ecosystem.”
For now, the Commerce Department’s order requires Anthropic to obtain a license before offering the models to any foreign person. Anthropic does not currently verify citizenship for its users, and it employs foreign nationals. The order signals that other AI companies may face similar scrutiny, and BIS could invoke additional powers if the Anthropic case is contested.
Aidan Gomez, CEO of Canadian AI developer Cohere, said the episode highlights the need for diversified AI supply chains. “If you rely on centralized large players, you are at risk of losing access,” he said at the VivaTech conference in Paris.
The directive has not yet been challenged in court, and the administration has not provided a detailed rationale. Anthropic’s technical and administrative teams are reportedly in talks with U.S. officials, but no resolution has been announced.
The situation underscores a broader trend of U.S. export‑control policy being applied to high‑tech software and AI. The policy shift could affect the availability of advanced models for researchers, businesses, and governments outside the United States, and it may prompt other countries to reassess their own AI export controls.
As the industry watches, the next steps will likely involve legal challenges, negotiations for licensing, and potential adjustments to export‑control frameworks to clarify the status of cloud‑based AI usage.
