The restrictions cover Mythos and Fable, Anthropic's most powerful frontier models. The White House has not confirmed the Semafor report's central China access allegation, and Anthropic has denied the government raised China during export control discussions.
The White House this week imposed export restrictions on Anthropic's two most powerful AI models, citing, in part, reported but unconfirmed fears that a China-linked group had already accessed the technology, according to a Semafor report summarized by The Verge.
The restrictions cover Anthropic's Mythos and Fable systems — the company's frontier-tier models — and amount to one of the first times the US government has treated a commercial AI product as a controlled technology. The specific concern, as reported by Semafor, is "distillation," a technique in which a less capable AI is trained to mimic a more advanced one, replicating its behavior without copying its underlying code.
Neither the White House nor Anthropic has confirmed the Semafor report. Anthropic told Semafor that a spokesperson said the government did not bring up China during its discussions around export controls. Anthropic did not respond to a request for comment from The Verge.
The only on-record US government statement about the export control is a June 14 X post from Trump advisor David Sacks, who described the action as triggered by an alleged jailbreak vulnerability rather than China access. "A highly credible trusted partner of both Anthropic and the USG who was testing Fable came forward with a jailbreak of those guardrails," Sacks wrote. "The Admin asked Dario to fix the jailbreak or de-deploy the model. Dario refused. In reaction, the Admin issued the export control. The Admin did this reluctantly."
Anthropic published an on-record statement contesting the government's characterization. "The government did not provide specific details of its national security concern," the company said. "Our understanding is that the government believes it has become aware of a method of bypassing, or 'jailbreaking' Fable 5." Anthropic said it reviewed a demonstration of the technique and found it "relatively simple" — vulnerabilities that "appear relatively simple" and are "available via other models, including GPT 5.5." The company added: "We have not even received a disclosure of a concerning non-universal potential jailbreak that led to a harmful result."
According to the Wall Street Journal, the export control directive was triggered in part by cybersecurity research from Amazon. CEO Andy Jassy shared findings with the White House showing that, through a series of prompts, Fable 5 could serve up information used in cyberattacks. Amazon has not responded to a request for comment.
Security researcher Katie Moussouris, founder of LutaSecurity, posted on Bluesky: "I've seen the paper. It's not a jailbreak."
Former Commerce Department official Kate Koren told the WSJ that the White House's dislike of Anthropic may have influenced the decision. Anthropic and the Trump administration have been at odds over the company's refusal to allow its AI for mass surveillance of Americans or lethal autonomous weapons; in February, Trump instructed federal agencies to stop using Anthropic's AI, and the Defense Department designated it a supply chain risk.
That gap between the policy rationale reported by Semafor and the government's public response is the story. US export controls are ordinarily preceded by a documented threat assessment, a rulemaking process, and a public record. In this case, the controlling narrative is a single second-hand report, a contested jailbreak claim, and the company's own characterization of its newest models as too dangerous to release.
What is established: the White House acted, and the action covers Anthropic's two most powerful models. What is reported, not confirmed: that a China-linked group accessed those models. What is on-record but contested: that a jailbreak of Fable's guardrails posed a national security risk sufficient to warrant an export control. What remains unanswered is why the policy rationale is sourced to a single report, and what threshold of evidence the US government now requires before treating a commercial AI system as a controlled technology.