Why the US restricted & then restored Anthropic’s most powerful AI models

The reversal caps an episode that began on June 12, when the Trump administration ordered Anthropic to deny the models to all foreign nationals on national security grounds. The order came at a time when Washington had been stepping up oversight of advanced AI models amid concerns that these could be misused by military intelligence in China, Russia or other countries of concern.

Since the company could not reliably separate foreign nationals from other users in real time, it had shut access to the models for everyone, Americans included.

Fable 5, released in early June, is a locked-down public version of Mythos 5, a model Anthropic had held back from wider public release over its unusually strong ability to spot vulnerabilities that hackers could exploit. Mythos 5 went only to a small set of companies and governments as part of what Anthropic calls Project Glasswing, allowing these select groups and entities to discover other potential dangerous uses or vulnerabilities in their own infrastructure.

Anthropic first restored limited access to Mythos 5 to a small set of ‘trusted’ US organisations on June 26, before the Commerce Department lifted the curbs on both models entirely on June 30.

The restrictions, though, rippled across the AI industry and raised questions over whether export controls can be used to regulate AI model access.

The company has now implemented a new safeguard that blocks this behaviour, it said. Any blocked request will instead be routed to its older Opus 4.8 model.

Also Read: Chinese AI company Z.ai targets Anthropic’s Claude Mythos

Anthropic said that although this may be frustrating for users, the tradeoff was made to ensure that Fable 5’s other capabilities remain widely available. The company also warned it was “probably impossible” to make any AI model fully robust to “jailbreaks”, the term used to refer to techniques that bypass safeguards.

“There will be many minor jailbreaks, some narrow harmful ones, and although no universal jailbreaks for Fable 5 have been discovered at the time of writing, expert safety researchers continue to red-team it,” the company said.

A letter to Anthropic from US commerce secretary Howard Lutnick said the company had agreed to work diligently with the government on protocols for Mythos, Fable and future models, and to inform the U.S. government of any malicious activity. Lutnick, however, said that the department “reserves the right to reevaluate the decisions made in this letter and the necessity of reimposing a license requirement, should circumstances change or should Anthropic fail to adhere to its commitments”.

Increased scrutiny of AI models began in June, after US President Donald Trump signed an executive order establishing a voluntary framework for AI developers to offer “covered frontier models” to the government for up to 30 days before releasing them to trusted partners.

Anthropic rival OpenAI also faced restrictions. It said last week that it had delayed a full public launch of GPT-5.6 at the US government’s request, limiting its access to a small group of vetted partners.

The episode marks a shift in how Washington now views AI polices. Export controls have traditionally been used by the US for hardware that powers these models, not the finished and publicly deployed software. That the government briefly withdrew access to a consumer product suggests this kind of intervention could become a new normal in a post-AI world.

Also Read: Why Anthropic has edge over OpenAI in IPO race

Isaac Harris, executive director of the Frontier Security Institute, a nonprofit focused on AI and national security, told Reuters there now appears to be a process for standards of the US models. But, he added, “There’s still a question mark as to how equivalently dangerous capabilities coming from China with less guardrails will be handled by the administration in the US market.”

Dean Ball, a former AI adviser to the Trump White House and the lead drafter of the administration’s 2025 AI Action Plan, had said on a New York Times podcast that Anthropic’s relationship with the US government dates to the summer of 2024.

At the time, the Joe Biden-era Pentagon and Anthropic had agreed to use Claude in classified settings, including intelligence analysis, but with two carve-outs: no domestic mass surveillance, and no fully autonomous lethal weapons, Ball had said on The Ezra Klein Show.

Also Read: What Claude’s Fable 5 withdrawal could mean & Anthropic’s long feud with Trump govt

The Trump administration expanded the contract in 2025 on the same terms, Ball said, before the relationship soured through the autumn and broke into the open early in 2026.

The decisive rupture, he said, came when Emil Michael, the under secretary of war for research and engineering, asked Anthropic to delete the clause barring the use of Claude to analyse bulk-collected commercial data, the capability at the heart of the mass-surveillance objection.

The fight was also unusually personal and political, Ball noted.

US defence secretary Pete Hegseth had written on X that Anthropic’s “true objective is unmistakable. To seize veto power over the operational decisions of the United States military”, which was “unacceptable”. Trump had also branded Anthropic a “radical left, woke company” and its staff as “Left-wing nut jobs”.

Matters came to a head in February this year, when Hegseth threatened to invoke the Defense Production Act and to designate Anthropic a “supply chain risk”.

That label is normally reserved for technology judged too dangerous to sit anywhere in the US military’s supply chain, and has been used against foreign firms such as China’s Huawei over espionage fears. The tag had never before been turned on an American company before.

Trump then announced on Truth Social that every federal agency would stop using Anthropic technology: “We don’t need it, we don’t want it, and will not do business with them again!” The US President also warned the company to “be helpful during this phase out period, or I will use the Full Power of the Presidency to make them comply, with major civil and criminal consequences to follow”.

Anthropic, whose Pentagon work fell under a contract worth $200 million, said being designated a supply chain risk “would both be legally unsound and set a dangerous precedent”, and added: “No amount of intimidation or punishment from the Department of War will change our position on mass domestic surveillance or fully autonomous weapons.”

The defence department ultimately signed a deal with OpenAI, which says it holds the same red lines as Anthropic. OpenAI chief Sam Altman told staff in a note seen by the BBC that his company would reject military uses “such as domestic surveillance and autonomous offensive weapons”.

Anthropic subsequently filed two federal lawsuits against the defence department on 9 March, alleging that the ‘supply chain risk’ label was applied improperly and amounts to retaliation for the company’s protected speech on AI safety, in violation of its due process rights.

“The Constitution does not allow the government to wield its enormous power to punish a company for its protected speech. No federal statute authorises the actions taken here,” Anthropic had argued, telling a court the designation would cause “irreparable harm” and put “hundreds of millions, or even multiple billions, of dollars in lost revenue” at risk.

The cases remain unresolved. A San Francisco federal judge granted a preliminary injunction barring enforcement of the ban on Claude, but the DC appeals court denied Anthropic an emergency stay on 8 April.

(With inputs from agencies; this article is an updated version of one published in June)