In Anthropic, CIOs thought they were buying into an ethical AI supplier that wouldn’t let its models be used autonomously in the military kill-chain or for mass surveillance. Now those customers find their access to Anthropic’s most powerful AI models can be turned off by the US government on what the company claims is a flimsy pretext.
The US government ordered Anthropic to prevent foreign nationals, including its own employees, from using its most powerful AI models the company said in a blog post on Friday. “The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance,” it said.
Anthropic said the US government had issued an export control directive suspending access to the models citing national security concerns. The letter ordering the suspension arrived at 5.21 p.m. Eastern Time on Friday, and did not provide specific details of the concern, it said.
“Our understanding is that the government believes it has become aware of a method of bypassing, or ‘jailbreaking’ Fable 5. We reviewed a demonstration of this specific technique being used to identify a small number of previously known, minor vulnerabilities,” Anthropic wrote.
It said it had been told of a potential narrow, non-universal way of “jailbreaking” the model by asking it to read a specific codebase and fix any software flaws. “We have reviewed a report that we believe is the basis of the government’s directive and validated that the level of capability displayed there is widely available from other models (including OpenAI’s GPT-5.5), and is used every day by the defenders who keep systems safe,” Anthropic wrote.
While the spirit of the export control order reinforces concerns about sovereign control of the IT systems companies and governments rely on, the letter of it — and Anthropic’s broad response — raises questions of identity management: A user’s nationality is not typically associated with their login credentials, and even if it were, there is no telling the nationality of the person at the keyboard once an account is connected. The decision will have rung alarm bells for European CIOs, said Matthew McDermott, director of AI at Access Partnership. “This is what it means to depend on US frontier models for critical workloads. It’s no longer enough to ask whether a model is accurate or secure. You also have to ask which government can order it to be switched off, and on what terms. Any organization that has wired core workflows, code‑assistants or cyber‑defence tools directly to a single US model has just discovered a new single point of failure,” he said.
Rik Turner, chief analyst at Omdia, agreed. “From the perspective of European organizations, the episode serves to further illustrate their continent’s relative impotence in the AI arms race. The majority of available frontier models are either US-owned and closed-source or Chinese and open-source,” he said. “Now the US government has shown it can turn off access to those models at the drop of a hat, it makes the case for AI sovereignty.”
Thomas Reuner, senior analyst at PAC, suggested that even in a market as young as that for AI services, there is an element of inertia behind buying systems that means Anthropic’s enterprise users are unlikely to jump ship. “If you’re already using Anthropic, it’s highly unlikely that you’re going to change. After all, if you start off as a Microsoft shop, you’re going to end as a Microsoft shop.”
The US government decision can even be seen as a vote of confidence in the power of Anthropic’s products, said Jason Holloway, managing director of QL Security, a UK-based AI security company, “It’s a great boost to Anthropic as it shows how ahead of the game they are.”
He said that no other company is as far down the line as Anthropic. “OpenAI does have its own model, but it’s not been on general release yet. But, when it does, it could find itself banned too,” he added.
Holloway expects the affair to blow over and for Anthropic to quickly restore access to Fable and Mythos: “It may be days, it may be a few weeks, but they’ll be back,” he said.