US administration gives Anthropic 90 minutes to restrict foreign access

The Trump administration ordered the suspension of Anthropic – one of the most advanced AI models, in a move that has stunned the industry and raised questions over how Washington plans to regulate it.

The Department of Commerce on June 12 imposed export controls on Anthropic’s newest models, Fable and Mythos, restricting access by foreign nationals and forcing the company to suspend both systems globally, Caliber.Az reports, citing the Financial Times.

According to a person familiar with the matter, Anthropic was given just 90 minutes to comply and was not provided with detailed explanations for the decision before the order was issued.

In response, the company has flown a senior technical team to Washington in an effort to manage the fallout.

The directive followed findings from Amazon researchers who identified a potential “jailbreak” in the Fable model — a method of bypassing safeguards to extract information about software vulnerabilities. Amazon shared its findings with Anthropic.

Amazon CEO Andy Jassy also discussed the issue with US officials, with sources saying he raised broader concerns about frontier AI capabilities rather than issues specific to Anthropic.

Amazon said it is common for governments to seek input on potential security risks but declined to disclose details of such discussions.

The Fable model had reportedly been approved for release by agencies within the Commerce Department just days before its public launch, according to people close to the matter.

Anthropic has pushed back against the decision, arguing that the vulnerability is not unique and can also be found in rival systems, including those developed by OpenAI.

Security experts say fixing such issues would require weeks of research and development, with no guarantee that similar vulnerabilities would not reappear.

The case is being seen as an early test of how the Trump administration intends to regulate increasingly powerful AI systems. Industry figures warn that using export controls to address model safety concerns could create uncertainty around the deployment of advanced AI tools.

“It is a pretty widely agreed upon fact that you cannot fully fix jailbreaks in these models, it’s a very inexact science” said Helen Toner, former OpenAI board member and interim executive director at Georgetown University’s Center for Security and Emerging Technology. She added that similar capabilities likely exist in other leading models.

The government’s restrictions do not currently apply to competing systems from companies such as OpenAI, despite reports that they demonstrate comparable capabilities.

The intervention is notable given Anthropic CEO Dario Amodei’s public warnings about AI safety risks. The company’s latest Mythos model has reportedly shown the ability to bypass certain cyber defenses, and had been developed in coordination with U.S. officials before its wider release.

The dispute also comes amid ongoing tensions between Anthropic and parts of the U.S. government, including disagreements over AI regulation and the use of AI in surveillance and autonomous weapons. The Pentagon has previously flagged Anthropic as a potential supply-chain risk.

The Trump administration has described the move as a national security measure unrelated to those broader disputes. A U.S. official suggested that addressing the jailbreak issue could resolve the standoff.

However, Defence Secretary Pete Hegseth linked the decision to past criticism of the company, writing on X that the Pentagon had previously kicked the company “out of our building — forever,” adding that events were proving that decision correct.

Despite the restrictions, US defence and government agencies are still reported to be using Anthropic’s advanced models.

The European Commission, which recently gained access to the company’s latest systems, said it was assessing the implications of the directive, warning that AI models bring both major benefits and serious cybersecurity risks.

“This is a shared challenge, not one confined to a single jurisdiction or company,” the Commission said, adding that any contingency measures should avoid discrimination against international partners.

Former OpenAI board member Katie Moussouris said the capabilities identified in Amazon’s report were not unusual and argued that similar functions exist in other major AI systems, including OpenAI’s GPT-5.5. She said the safeguards in question were functioning as intended.

“The administration got it wrong,” she said.

By Bakhtiyar Abbasov