US bans foreign nationals from accessing Anthropic's top AI models over jailbreak fears

By Jung Yoon-hee, The Herald Business

The US administration has imposed a blanket ban on foreign nationals accessing Anthropic's top-tier AI models, including Mythos.

Anthropic said Thursday that the US government had issued an export control directive, under guidance from national security authorities, ordering the immediate suspension of all foreign nationals' access to its Fable 5 and Mythos 5 models.

The directive bars not only access to Mythos from abroad but also access by foreign nationals physically present inside the United States. Anthropic's own foreign national employees are also subject to the restriction.

Anthropic said it received the government directive at 5:21 p.m. Eastern time that day and immediately suspended Fable 5 and Mythos 5 service for all customers to comply with the order.

However, Anthropic said the letter it received did not specify the national security concerns behind the decision.

The company said it believes the government may have concluded that users can bypass Fable 5's safety guardrails — designed to prevent malicious hacking, the creation of biochemical weapons and other harmful uses — through a technique known as jailbreaking.

Anthropic said it considers the situation a matter of misunderstanding and said it would work to restore service as quickly as possible.

The company added that the jailbreaking method the government appears to have identified is widely used across other models, including OpenAI's GPT-5.5, and is employed daily by security professionals. Anthropic said it believes the government should have authority to block unsafe AI deployments through transparent, fair and technically grounded legal processes, but said this action does not meet those standards.

The US Commerce Department decided to act after a company claimed it had breached Mythos's security barriers, raising national security concerns, Axios reported Thursday, citing government sources.

The government had earlier attempted to halt the release of Anthropic's models but failed, after which it sent the export control notice, sources told Axios. The access restrictions are expected to remain in place for several weeks until US government security frameworks are strengthened.

Mythos drew wide alarm in the global security community — a reaction dubbed the "Mythos shock" — after it emerged that the model possesses expert-level vulnerability detection capabilities.

Anthropic first disclosed the development of Mythos in April but, given its potential risks, shared it only with a limited set of institutions — including government agencies, technology companies and financial firms — before publicly releasing Fable 5, a version with safety guardrails applied, on June 9.