Anthropic received a directive from the U.S. government on a Friday evening in June 2026. It ordered the company to block access to its two most advanced models, Fable 5 and Mythos 5, for any foreign national anywhere in the world. The company could not separate those users from its broader audience in real time. So it shut the models down for everyone.
Other Claude versions stayed available. But the flagship systems went dark. Overnight. No public explanation accompanied the order at first. National security authorities were cited. Details emerged later. The concern centered on a demonstrated jailbreak method. Officials worried the vulnerability could let adversaries, possibly linked to foreign military intelligence services, divert the models for harmful ends.
Commerce Secretary Howard Lutnick signed the letter. It invoked powers under the 2018 Export Control Reform Act for the first time on an AI model in this manner. The directive required a license for any transfer or access by foreign nationals, even those inside the United States. Failure to comply carried the threat of criminal and civil penalties. (Reuters)
Anthropic had tested Fable 5 rigorously with U.S. and U.K. agencies beforehand. The company maintained that the flaws were minor. They existed in other publicly available models too. A full bypass of all safeguards did not occur. Still, the government acted. Anthropic complied by disabling global access. It called the move abrupt. Executives dispatched technical staff to Washington that weekend. They held calls with officials including Lutnick and National Cyber Director Sean Cairncross. The goal was to resolve the restrictions quickly.
More than 80 cybersecurity executives signed an open letter days later. They urged the administration to lift the controls. “This action has taken the best models away from defenders, created market uncertainty, and risked America’s AI leadership without any real risk to justify it,” the letter stated. (The Wall Street Journal)
The incident highlights a deeper tension in how Washington approaches frontier AI.
Export controls long targeted hardware. Chips. Tools. Software code. Model weights drew attention under the prior administration. The Trump team walked back some of those ideas. Yet here the controls landed on access itself. Users query the models through hosted services. No weights change hands. No download occurs. The “export” concept stretches. Remote cloud access has long been a recognized gap in the rules. Congress has considered legislation to close it.
Experts called the application unprecedented. “To my knowledge, this is the first time US export controls have been used to control access to an AI model in this way,” said Hanna Dohmen, senior research analyst at Georgetown University’s Center for Security and Emerging Technology. The legal basis remains partly opaque. Andrew Reddie, professor at UC Berkeley’s Goldman School of Public Policy, described the area as unsettled. “To say that this is an unsettled area of export control rule-making would be an understatement.” Successive administrations have sent mixed signals about developer responsibilities. That leaves companies guessing. (The Verge)
Anthropic itself once pushed for strong compute controls. In April 2025 it submitted recommendations supporting the Department of Commerce’s Framework for Artificial Intelligence Diffusion. The company argued that maintaining America’s advantage in advanced semiconductors served national security and economic interests. Chinese firms had stockpiled chips ahead of deadlines. Delays would only weaken enforcement. Powerful AI systems could arrive by 2027. The first nation to build them would gain a decisive edge. (Anthropic)
Now the controls hit the models themselves. The company’s earlier advocacy for pre-deployment review boomeranged in a way. It sought transparent thresholds and process. The letter it received offered neither. Anthropic described the concerns as a misunderstanding. It worked to restore access. David Sacks, the president’s AI policy advisor, called the issue easily resolvable and separate from other disputes. Yet the precedent lingers. The government demonstrated it can pull frontier models offline worldwide on short notice. Charges stay partly unseen. Companies cannot fully argue against them.
Analysts point out the self-defeating aspects. If the capability poses a novel cybersecurity threat, rapid action before full disclosure might make sense. But the vulnerability appeared routine. Other models could replicate similar behaviors. Shutting down one provider does not erase the knowledge. It simply shifts usage elsewhere. Defenders lose access. Uncertainty spreads across the industry. Investment calculations change. International partners question reliance on American systems. Europe weighs its sovereignty gap. (Doug Levin on Substack)
Relations between Anthropic and the administration had already frayed. The company refused certain military uses of its technology for surveillance or autonomous weapons. That led to blacklisting in some contexts. The export order amplified existing friction. Yet meetings continued. Technical staff engaged daily with Commerce officials. Dario Amodei and Lutnick both planned to attend G7 gatherings. A path toward resolution existed. Still, the episode exposed cracks.
Governance of advanced AI cannot rest on ad hoc directives. Companies need clear standards. What level of jailbreak resistance suffices? Which capabilities trigger intervention? How does one verify nationality across millions of users without broad shutdowns? These questions demand answers before the next generation of models arrives. Without them, American leadership erodes. Competitors abroad gain from the hesitation. The very controls meant to protect advantage may hand it away.
The Fable and Mythos shutdown lasted days. Access began returning in stages after negotiations. But the damage to trust and predictability will take longer to repair. Industry insiders watched closely. They saw a blunt instrument applied to a complex problem. Future releases will carry new caution. Boards will weigh regulatory risk higher. Innovation may slow in the name of compliance. Or it may accelerate elsewhere, outside U.S. jurisdiction.
This was no isolated enforcement action. It marked a shift. Washington showed willingness to treat hosted AI access as an export. The implications stretch beyond one startup. Every frontier lab now operates under the shadow of similar potential orders. The race for superiority continues. But the rules of engagement just grew more uncertain. And the stakes keep rising.
WebProNews is an iEntry Publication