Anthropic said on Friday night that the U.S. government had ordered the company to suspend access for all foreign nationals to its Fable 5 and Mythos 5 artificial intelligence models, citing national security concerns.
The company revealed the order in a social media post, saying that it “must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance.” Foreign nationals who work for Anthropic are also affected by the suspension.
Anthropic unveiled Claude Mythos in April, but said it could not share the A.I. system with the public because it could become a powerful tool for hackers looking for ways to break into computer networks.
The new A.I. system quickly became a security concern in Washington and Silicon Valley because it could represent what experts had worried about for several years:that A.I. represented a new and far more dangerous threat to computer networks.
Anthropic shared the Claude A.I. system with about 40 organizations that maintain critical computer infrastructure so they could use the system to patch security vulnerabilities before hackers exploited them.
The order to limit access to the new A.I. models came from the Commerce Department, said a person familiar with it who spoke on the condition of anonymity. It was unclear how long the restrictions will last.
Anthropic declined to comment beyond its post, and the White House did not immediately respond to a request for comment.
Reactions from researchers who have had access to Mythos have varied. Some agreed that Mythos was a new and alarming threat. But others said that it was more evolutionary than revolutionary, and that they could use it to protect their networks just as hackers could use it to break into them.
In its brief statement on Friday, Anthropic seemed to indicate that it considered the government’s decision an overreaction. “We apologize for this disruption to our customers. We believe this is a misunderstanding and are working to restore access as soon as possible,” the company said.
After Anthropic unveiled Mythos, the White House started considering new oversights of A.I. systems. Last week, in a significant shift in policy for an administration that had taken a very hands-off approach to A.I., President Trump signed an executive order that asked companies to voluntarily give the government oversight of new systems before releasing them to the public.
Earlier this month, Anthropic said it would share the system with about 150 other organizations. Then, on Tuesday, the company unveiled a straitjacketed version of Mythos called Claude Fable 5.
The Fable system, the company said, included additional guardrails designed to block responses in areas like cybersecurity, biology and other sensitive topics. The guardrails may prevent malicious hackers from using the technology to attack computer networks. But the guardrails may also prevent companies and cybersecurity experts from using the system to defend against attacks.
Most queries that it deemed as risky, Anthropic said, would instead be shuttled to an older A.I. system called Opus 4.8.
Anthropic’s Opus 4.8 and other older A.I. models were not affected by the government’s order on Friday, the company said.
Cybersecurity experts have disagreed on whether Anthropic was right to limit the release of Mythos, pointing out that this kind of system can be used to both attack and defend computer networks. Other companies, most notably the Anthropic rival OpenAI, offer similar technology.
Former U.S. cybersecurity and technology policy officials said they were surprised by the news, calling it highly unusual and out of step with other moves by the Trump administration, such as allowing some advanced A.I. chips to be sold to China.
“I have no words,” Dean Ball, a senior fellow at the Foundation for American Innovation and a former adviser for A.I. under Mr. Trump, said in a social media post. He added that the decision was “baffling.”
A week after Anthropic unveiled Mythos, OpenAI said that it would also limit the release of its latest system, called GPT-Cyber 5.4. But OpenAI shared its technology with a larger group, initially sharing the system with hundreds of experts before opening it up to thousands.