Anthropic shuts down Mythos access after sweeping U.S. order

Anthropic PBC disabled access to its most advanced AI models, including Mythos, following an unprecedented order from the Trump administration to keep the technology out of the hands of all foreign nationals.

In the statement made by the company, it was stated that the US government told Anthropic to suspend access to the Fable 5 and Mythos 5 models by any foreign national “inside or outside the United States”, citing national security concerns.

A US official confirmed that the Department of Commerce sent the letter. The model developer has since shut down all customers’ access to both systems to ensure compatibility.

Never before has the US government taken such sweeping measures to rein in foreign access to frontier AI models developed by an American company. The Trump and Biden administrations have limited access to other important technologies such as semiconductors and supercomputers abroad, and some have debated the merits of blocking access to artificial intelligence models. But restrictions on the software itself have raised constitutional and commercial concerns.

Anthropic said it believed the U.S. government issued the order after discovering that it was possible to “jailbreak,” or bypass the guardrails, of Fable 5, a recently released version of Mythos that prevented the company from performing cybersecurity duties.

“We disagree that a small finding of potential jailbreak should be a reason to recall a commercial model that has been applied to hundreds of millions of people,” Anthropic said in a post on its website. “We believe that if this standard were implemented industry-wide, it would essentially halt all new model deployments for all lead model providers.”

According to a report in the Wall Street Journal, researchers at Amazon.com Inc. conducted a jailbreak study that revealed some vulnerabilities in Anthropic’s model.

Amazon and the U.S. government were in contact about the vulnerability before the controls were implemented, said people familiar with the matter who were granted anonymity to discuss sensitive conversations. Amazon Chief Executive Officer Andy Jassy was involved in those purchases, one of the sources said. Information had previously reported that Jassy had raised concerns with senior US officials.

An Amazon spokesman said it was not unusual for governments to consult with the company about security risks, but declined to share details of such discussions.

The government’s move to so broadly restrict access to a range of AI models in the name of national security threatens to set a precedent for all major AI model developers, including OpenAI, Alphabet Inc.’s Google and Meta Platforms Inc. Nvidia Corp. Industry leaders such as Chief Executive Jensen Huang and OpenAI CEO Sam Altman have in the past encouraged the U.S. government to instead encourage worldwide adoption of American AI systems and maintain the country’s leadership.

Nvidia Corp. “This is a huge wake-up call for anyone who was naive and maybe hoping that this power wouldn’t be used,” Aidan Gomez, co-founder of backed artificial intelligence startup Cohere Inc., said in an interview on Saturday. “Nobody can deny it now.”

Anthropic said it received the government order at 5:21 p.m. New York time on Friday. The end-of-day guidance contradicts earlier statements, as well as an executive order recently signed by President Trump that suggested the administration would not pursue a licensing regime for model reviews.

Friday’s directive also threatens to escalate long-standing tensions between Anthropic and some within the Trump administration. Earlier this year, the artificial intelligence developer clashed with the Pentagon over the use of its technology for military and surveillance purposes. As a result of the explosion, the administration declared the company a US supply chain risk and ordered US agencies to phase out the use of its products.

Private company Anthropic, which has long positioned itself as a more responsible AI developer, first rolled out its Mythos model to a very limited group of companies and institutions in April, warning that its ability to find cybersecurity vulnerabilities made it too risky to deploy it more widely.

There were signs that the limited release was an attempt to ease tensions between Anthropic and the Trump administration: In April, the US government was preparing to make a version of Mythos available to major federal agencies, Bloomberg previously reported.

Mythos has also stepped up the Trump administration’s efforts on AI policy, which includes a recent executive order calling for voluntary model review. This order clearly said that nothing in it should be construed as creating a compulsory licensing regime.

David Sacks, Trump’s former AI czar and current co-chairman of the President’s Council of Science and Technology Advisers, said Anthropic refused to fix the jailbreak of guards in the Fable model.

“The admin’s hope is that Anthropic will fix the security issue, export control will be removed, and Fable will go back to general release,” he wrote in a post on X. “The executive wants all of this to happen as quickly as possible. Frankly, it is surprising that Anthropic does not want to comply with security requests that it has previously said is its highest priority.”

The government’s latest crackdown clashes with a race among U.S. developers to deliver the most advanced AI models and prove to investors that the technology can turn a profit. Both OpenAI and Anthropic are seeking to go public as soon as this year, following SpaceX’s own historic IPO.

The rush to deliver cutting-edge AI models spurred Anthropic to publish a lengthy blog earlier this month; He called for the creation of a system where governments and artificial intelligence developers would collectively decide when to slow down work on the technology to eliminate the risks it could pose.

“Having the option to show or temporarily pause potentially dangerous AI work would be good for the world,” the company said in a post at the time. The company said artificial intelligence is advancing to the point where the technology could make human work thousands of times more efficient or even replace it, creating a new set of risks.

The European Union’s executive arm said it was evaluating Anthropic’s statement and continues to talk with allies about potential risks and cybersecurity concerns related to powerful new artificial intelligence models. The European Commission added that recent developments underline Europe’s need for technological sovereignty.

“As someone who works in the field, I wasn’t very excited to see this,” Cohere’s Gomez said. “I don’t think it’s a partnership, I don’t think it’s the right thing to do for the broader technological alliances that have developed over the last 80 years.”

Eastland and Lowenkron write for Bloomberg. With assistance from Shirin Ghaffary, Yi Wei Wong, Gian Volpicelli, Spencer Soper, and Thomas Seal.