White House forces Anthropic to pull AI models after suspected Chinese breach

The Trump administration moved fast last week to impose export restrictions on two of Anthropic's newest artificial intelligence models after security concerns, including suspicions that a China-linked group had gained access to one of them, triggered a series of White House meetings and a public standoff between the AI company and senior officials.

Anthropic blocked public access to its Fable 5 and Mythos 5 models on Friday, responding to a directive from Commerce Secretary Howard Lutnick. The company had released both models publicly just days earlier, on Tuesday. By the end of the week, the models were gone, pulled from every user, domestic and foreign alike.

The sequence of events, first reported by the Washington Examiner, paints a picture of a company that talked a big game about AI safety but resisted when the federal government asked it to act on a specific, concrete threat.

Amazon researchers found the flaw

The crisis began when researchers at Amazon managed to jailbreak the Mythos 5 model, bypassing its built-in security features to produce outputs that should have been blocked. The Wall Street Journal reported that Amazon's chief executive, Andy Jass, told U.S. officials about the discovery, including Treasury Secretary Scott Bessent.

That revelation set off a chain reaction inside the administration. On Friday, White House chief of staff Susie Wiles and President Donald Trump were involved in a series of meetings at the White House to address the situation, the Wall Street Journal reported.

The concern went beyond a simple software bug. Semafor reported Saturday that the White House imposed restrictions partly over suspicions that a China-linked group had accessed Mythos 5, citing a person familiar with the matter. The export restrictions initially prohibited access by customers outside the United States and by foreign nationals within the country.

But Anthropic did not simply restrict foreign access. It went further, removing access to Fable 5 and Mythos 5 for all users, including Americans, in response to Lutnick's directive.

Anthropic pushed back before complying

White House AI czar David Sacks laid out the administration's side of the dispute in a post on X Saturday. He said the administration asked Anthropic CEO Dario Amodei to either fix the jailbreak or pull the model. Amodei refused, Sacks said.

Sacks wrote that the restriction was put in place "reluctantly."

"In the past, Anthropic has always said that safety must be top priority and taken super seriously. In this case, Anthropic prioritized the continued offering of the consumer model over safety."

That is a sharp charge from a senior official, and it cuts against the brand identity Anthropic has built for itself. The company has positioned itself as the safety-first alternative in the AI race, the responsible actor in a field of reckless competitors. Sacks essentially accused the company of abandoning that posture the moment it became commercially inconvenient.

The administration's willingness to use executive authority to enforce AI security fits a broader pattern. The White House has moved aggressively to reshape federal institutions and assert direct control over agencies and industries it views as insufficiently responsive to national security priorities.

Anthropic's defense

Anthropic did not go quietly. In a public statement, the company pushed back on the premise that a single jailbreak vulnerability should justify pulling a product used by a massive audience.

"That the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people. If this standard were applied across the industry, we believe it would essentially halt all new model deployments for all frontier model providers."

That argument has a surface logic to it. Every major AI model has been jailbroken at some point. Security researchers find new exploits constantly. If a single vulnerability meant an automatic shutdown, no AI model would stay online for long.

But the argument sidesteps the specific allegation at the center of this dispute: that a China-linked group may have already exploited the flaw. A theoretical vulnerability is one thing. A suspected active breach by a foreign adversary is something else entirely.

Katie Moussouris, CEO of cybersecurity firm Luta Security, reviewed a report shared by Anthropic and said the jailbreak could "generate output that could be used to find vulnerabilities and generate test scripts that could be turned into proof of concept code." That is not a trivial finding.

The Pentagon connection

Secretary of War Pete Hegseth weighed in Saturday night with his own post on X, referencing a prior standoff between the Pentagon and Anthropic. Hegseth said "every passing day proves" that the Pentagon's earlier decision to end its government contract with Anthropic was the right call.

The details of that prior dispute are not fully clear from the available reporting. But Hegseth's comment suggests the administration views the current episode as confirmation of a pattern, a company that talks about safety while resisting the kind of accountability the government expects from firms handling sensitive technology.

That tension between Silicon Valley's self-regulation rhetoric and the government's enforcement authority is only going to intensify. The legal and constitutional boundaries of executive power in technology policy remain a live question, with major cases working through the courts on related questions of presidential authority.

Mythos 5 and Project Glasswing

Mythos 5 first drew national attention in early April, when Anthropic released it on a limited basis through a program called Project Glasswing. The company described the initiative as an effort to secure "the world's most critical software," granting access only to a select group of companies.

That controlled rollout made the subsequent decision to release Mythos 5 publicly, alongside Fable 5, all the more striking. The model went from a carefully guarded tool available to vetted partners to a consumer product in a matter of weeks. And within days of the public launch, the administration was scrambling to contain the fallout from a suspected foreign breach.

Anthropic filed to go public earlier this month, adding a financial dimension to the dispute. A company preparing for an IPO has powerful incentives to keep its flagship products available and its user numbers growing. Pulling two major models from public access, even temporarily, is the kind of move that complicates a market debut.

The administration's restructuring of the federal workforce has touched nearly every corner of government, from mass departures of federal lawyers to the removal of advisory board members. The Anthropic episode extends that pattern into the private sector, with the White House signaling that companies handling frontier AI technology will face real consequences if they fail to address security threats promptly.

Amazon's role

Amazon's involvement adds another layer. The company's researchers were the ones who discovered the Mythos 5 jailbreak. An Amazon spokesperson acknowledged the company's interactions with the government but declined to provide specifics.

"As a leading cloud provider that serves a large number of private and public sector customers, it's not uncommon for governments to seek our counsel on potential security risks. When they occur, we don't share details of these discussions."

Amazon is a major investor in Anthropic and hosts its models on its cloud infrastructure. The fact that Amazon's own researchers found and reported the vulnerability, rather than covering for a business partner, suggests the company took the security concern seriously enough to escalate it through official channels.

The broader question of how the federal government manages AI security is far from settled. The administration has moved to reshape science and technology oversight boards, and the Anthropic case may accelerate calls for more formal regulatory frameworks around frontier AI models with national security implications.

What remains unanswered

Several important questions remain open. What specific evidence supports the suspicion that a China-linked group accessed Mythos 5? What legal authority did the Commerce Department invoke to impose the export restrictions? And what exactly did the Pentagon's earlier standoff with Anthropic involve?

The full text of Lutnick's directive has not been made public. Nor has the administration disclosed the intelligence or technical evidence behind the suspected Chinese access. Those details matter, both for evaluating the severity of the threat and for understanding the legal precedent being set.

What is clear is that the administration moved within days of the public release to shut down access, and that Anthropic initially resisted before complying. The company that built its reputation on putting safety first had to be told, twice, to act on a security threat that may have involved America's chief geopolitical rival.

A company headed for an IPO that won't pull a compromised model when the government says a foreign adversary may have gotten in is not a company that gets to lecture the rest of the industry about safety.