A late-night call landed at Anthropic on Friday and the room went quiet. I heard the detail that the company had roughly 90 minutes to pull two of its flagship models offline. You can feel how quickly a technical problem becomes a geopolitical one — the call was a live wire in Washington.
I’ll walk you through what’s verifiable, what reporters are piecing together, and what the near-term stakes look like for you if you follow the AI industry. Read to the end for a reminder of what else President Trump is up to today.
A late-hour demand arrived: Anthropic was told to shut models down
A late Friday call reportedly told Anthropic that keeping Claude Fable 5 and Claude Mythos 5 online posed a “national security threat.” Axios reports that the White House warned the company and an export control order followed, instructing Anthropic not to allow the models to be used by non-U.S. nationals.
Those two models are part of Anthropic’s Mythos-class family, descended from the controversial Claude Mythos Preview the company discussed in April. Anthropic had positioned Fable 5 and Mythos 5 as safer, guardrailed variants and briefly offered them to paid Claude users from Tuesday through Friday. According to reports, hours after the call the company disabled them.
Why did the White House order Anthropic to take the models offline?
Reporting from Axios, Reuters and Semafor points to demonstrations of jailbreaks and concerns about access by a China-linked group. Amazon and at least five other companies notified the White House that they could bypass the models’ safeguards, per Axios and Reuters, prompting the administration to act with export controls and an urgent demand.
Companies raised alarms: Amazon and others flagged potential jailbreaks
Private-sector demonstrations fed the federal response. Sources told reporters that Amazon was among the firms that signaled proof-of-concept jailbreaks to the White House.
Those signals matter because they move this from an internal safety failure to an operational, cross-industry vulnerability. When cloud providers or large platform operators flag a risk, the options narrow fast: patch, pull, or be regulated into submission. Anthropic’s models are hosted on major infrastructure and the chain of custody and control quickly becomes a national-security conversation.
Can non-U.S. nationals still use Claude Fable 5 or Mythos 5?
No — the export control order specifically targeted access by non-U.S. nationals, which effectively halted public availability. The models were taken offline after the White House letter arrived, and Anthropic reportedly had only a short window to comply.
Anthropic’s leadership went to Washington on Sunday to try to resolve the crisis
Senior Anthropic figures flew to D.C. for talks with the Trump Administration on Sunday, according to sources cited by Axios. The White House, per reporting, was dissatisfied with Anthropic’s initial responses and wanted more serious engagement.
It’s rare to see negotiations move this fast, and rare that an AI product’s operational status becomes a White House communications incident. The company now finds itself managing corporate legal risk, partner relations with Amazon and others, and an intensifying public regulatory gaze that includes media such as Reuters and Semafor.
What happens next for Anthropic and the wider AI industry?
Expect three immediate threads: an operational forensic to prove the jailbreaks are fixed, legal and export-control work to demonstrate access controls are airtight, and a dialogue with cloud providers (including Amazon Web Services) about hosting practices. Industry figures from Anthropic and its peers — and regulatory attention that has trended toward OpenAI and other major labs — will shape the next moves.
Context matter: the models were public for a brief run
Real-world observation: People with paid Claude accounts had access to these models for only a few days. That window is one reason the story escalated so quickly.
The short release was supposed to be a controlled test of Mythos-class guardrails. Instead, companies demonstrated jailbreaks and the federal government treated the situation as an export-control and national-security issue. That shift is a reminder that releasing advanced models carries immediate geopolitical consequences.
Brands and reporters in play: Anthropic, Claude Fable 5, Claude Mythos 5, Claude Mythos Preview, Amazon/AWS, the White House, Axios, Reuters, Semafor — all are part of the public record on this. You should watch statements from Anthropic and any follow-ups from the White House and Amazon for the clearest updates.
There’s an obvious tension here: companies want to field powerful models quickly while governments want to limit potential misuse. Think of the negotiations ahead as a chess match over who controls access to powerful code.
Finally — and this is the small, theatrical detail — Sunday is President Trump’s 80th birthday, and a UFC octagon has been set up on the White House lawn for a scheduled match. Does an urgent AI security negotiation change the optics of a day already scripted for spectacle?