US Export Clampdown on Anthropic’s AI Models Shakes the Industry

Former Trump AI adviser Dean Ball reacted to the news with just, "I have no words," a line that quickly became one of the most quoted responses to the controversy. The reaction followed reports that the U.S. Commerce Department ordered Anthropic to suspend access to its most advanced AI models for all foreign nationals, a move that stunned much of the AI sector.

Anthropic said on Friday it had been told to block access to its models Mythos 5 and Fable 5 for every foreign citizen, whether outside the United States or living and working inside it, including the company’s own foreign employees. To comply, Anthropic said it would have to shut access to the models for all users worldwide until further notice. The case has triggered broad debate in Silicon Valley and the global cyber community because U.S. export controls have traditionally targeted chips, GPUs and semiconductor equipment, not the model itself.

Australian research firm Isaacus warned that this may be the first time the United States has imposed export limits on a large language model rather than hardware. It said the decision could affect not only adversaries but also close U.S. allies, including Britain, Canada, Australia and New Zealand, and that any app or service relying on an American model could now lose access overnight. The firm asked: "Whoever depends on AI for critical tasks must ask what happens when that intelligence simply disappears overnight."

U.S. reports say Commerce Secretary Howard Lutnick issued the emergency order after Amazon Web Services, Anthropic’s main investor and partner, identified a critical weakness and demonstrated it to the government. The concern is that jailbreak techniques could bypass Fable 5’s safeguards and expose Mythos 5’s cyber capabilities, enabling hostile actors to write exploits against critical U.S. infrastructure. Anthropic rejected that account, saying it was not shown evidence of a major security flaw or a broad bypass method, only relatively simple software bugs also found by other public models. The company said it had spent months on security testing with U.S. and British officials, internal researchers and outside security groups, and none found a reliable way to defeat the protections. The move has also renewed debate over Anthropic’s own safety design, its 30-day storage policy for usage data and prompts, and whether AI models should be treated as strategic assets like advanced chips. Some officials welcomed the step, including Pentagon information chief Kirsten Davies, who said national security outweighs business interests.