U.S. curbs pushed Anthropic to take major AI models offline

The Trump administration’s emergency export restrictions forced Anthropic to remove its Fable 5 and Mythos 5 AI models from service, Fortune reported, marking a new and aggressive step in U.S. oversight of advanced AI. The decision matters because it placed national security controls on publicly available frontier AI systems after concerns that their cyber abilities could be misused.

According to Fortune, the immediate trigger was a security test by Amazon researchers on Fable 5, a version of Anthropic’s Mythos model that the company had described as safer for broader release. The researchers found a jailbreak that they said could let users get around safeguards and obtain information useful in a cyberattack, Fortune reported.

Amazon told Anthropic about the finding, according to Fortune. The issue escalated when Amazon CEO Andy Jassy, during a previously scheduled June 11 call with White House officials on another matter, raised the vulnerability and was encouraged to alert Treasury Secretary Scott Bessent, Fortune reported, citing two people familiar with Amazon.

Jassy spoke with Bessent the same day and described the Fable issue, while also raising broader concerns about cyber capabilities across frontier AI models, according to Fortune. Bessent has been involved in the administration’s response because officials feared Mythos-enabled cyberattacks could threaten the financial system, Fortune reported.

The next day, the administration told Anthropic and CEO Dario Amodei to either fix the jailbreak or pull the models, according to Fortune. A person familiar with Anthropic told Fortune the company had not received earlier notice that the government viewed Fable’s export as a national security threat.

The Commerce Department then gave Anthropic a 90-minute deadline and imposed export controls that barred access by foreign nationals, Fortune reported. Because U.S. “deemed export” rules can treat release to a foreign national as an export, Anthropic viewed the order as a practical ban on the models, including for non-U.S. employees, according to Fortune.

Commerce Secretary Howard Lutnick’s formal letter arrived after 5 p.m. Eastern on Friday, Fortune reported. Bloomberg News, which obtained the directive, reported that Anthropic would need government approval before making Fable 5 and Mythos 5 available abroad or to foreign nationals anywhere, and that violations could bring civil and criminal penalties.

Anthropic took both models offline by 10 p.m. Friday, Fortune reported. In a company blog post quoted by Fortune, Anthropic said it disagreed that “a narrow potential jailbreak” justified recalling a commercial model used by hundreds of millions of people, and warned that applying that standard across the field could stop new frontier model launches.

David Sacks, a venture capitalist advising the Trump administration, wrote on X that the administration viewed the issue as serious but fixable, Fortune reported. By Tuesday, talks between Anthropic and government officials had ended without lifting the controls, according to Fortune.

The dispute followed earlier tension between Anthropic and the administration. Fortune reported that the government had already clashed with the company over Pentagon use of Anthropic models, after Anthropic sought limits on uses involving autonomous weapons and domestic mass surveillance. The administration later told federal agencies to stop using Anthropic models and labeled the company a supply chain risk, a designation Anthropic has challenged in court, Fortune reported.

Anthropic had also expanded access to Mythos through its Glasswing program, Fortune reported. The Washington Post reported that after the administration approved a list of 111 trusted organizations, Anthropic later added about 50 more, including a South Korean telecom company that U.S. officials suspected of China ties. SK Telecom said there was no basis for linking it to China, according to Fortune.

Cybersecurity and legal experts questioned the government’s move. ThreatLocker CEO Danny Jenkins told Fortune that restrictions on technology development risk leaving the U.S. behind, while Charlie Bullock of LawAI told Fortune the export control letter appeared to rest on “very shaky legal ground.”

Politico reported Thursday that Anthropic and the U.S. government were discussing a deal to restore access and create a shared framework for judging when jailbreaks pose serious danger. Chris Cauri, Anthropic’s managing director of international, told reporters in Seoul that the company was confident the models would become available again in coming days, Fortune reported.

This story draws on original reporting from Fortune.