US Opens Security Vetting of Frontier AI Models as OpenAI and Anthropic Restrict New Releases

ChatGPT maker OpenAI said Friday it is holding back the public release of its newest model at the request of the Trump administration, the latest move in an unprecedented federal review of advanced AI systems for security risks. Rival Anthropic, maker of the Claude chatbot, announced hours later that the administration had approved a limited return of one of its most capable models. The Associated Press reported the developments on June 26.

OpenAI said its new product, called GPT-5.6 "Sol," will be available only to customers the administration has approved. The company described the arrangement as a temporary step on the "path to broader availability in the coming weeks." Anthropic said the Commerce Department had effectively banned its Fable 5 and Mythos 5 models days after their unveiling; on Friday the government lifted the block on Mythos 5, letting Anthropic redeploy it to "a small group of cyber defenders and infrastructure providers."

Two companies, two different paths

Both firms said their strongest new models would reach only tight circles of trusted partners. OpenAI has not named the roughly 20 customers cleared to use Sol so far. Anthropic pitched Fable as a safer sibling to Mythos, but pulled it offline to comply with a directive that barred foreign nationals from using the systems.

The White House said it continues to work with frontier labs on the problems that come with scaling the technology. The statements mark the first time two major US model builders have publicly tied a product launch to a government clearance process.

The order that started it

Trump signed an executive order on AI oversight in June. It set up a framework for federal agencies to review the national security risks of the most advanced AI systems for up to 30 days before they reach the public. The order calls participation voluntary, and officials have yet to finish building the review machinery.

Some of the president's allies have pointed at Anthropic and its chief executive, Dario Amodei, as the reason the scrutiny exists. David Sacks, who helps lead the administration's council of science and technology advisers, said on a recent podcast that Amodei "came to Washington a few months ago, back in April, and basically said that he had created a cyber weapon called Mythos." Sacks added that the claim "spiked the cortisol level" in the capital and carried "some truth" because the model showed advanced cyber skills.

Why officials fear the models

Anthropic warned earlier this year that Mythos was unusually good at finding software flaws. Officials worry that same skill could be turned around by malicious hackers to breach critical computer networks. OpenAI pushed back on the comparison for its own model. The company said Sol is "better at helping people find and fix vulnerabilities" than at carrying out attacks, and that it does not cross OpenAI's internal risk line.

OpenAI still acknowledged gaps in what it knows. "That uncertainty, along with the model's broader step change in capabilities, is why we are pairing the model's increased capabilities with stronger safeguards and a phased release," the company said. The phrasing points to a concern that a model safe on its own could cause harm once combined with other tools.

Lawmakers and researchers push back

Critics say the process lacks any legal footing. US Representative Lori Trahan, a Massachusetts Democrat who co-wrote a bipartisan bill to regulate AI, said she is troubled that "the Trump administration is deciding company by company who gets access to the newest AI model. No law. No process. No oversight. Just appointees in Washington deciding who's in and who's out."

A broad set of security specialists has made the same point about the Anthropic order. Alex Stamos, a Stanford cybersecurity researcher and former chief security officer at Meta, said on a call with reporters that "pretty much nobody in the cybersecurity industry believes that there's any factual basis for this action." Stamos reviewed an analysis prepared by Amazon, Anthropic's main cloud partner, and said he found no risks absent from other models already on the market, including ones built in China. "If the administration is honest about wanting the United States to beat China in this race, then this is about the dumbest thing they could possibly do," he said.

The dispute touches the physical layer of the business as well. The servers that train and serve these models sit in data centers that have drawn their own scrutiny over water, power, and siting. Our coverage of cloud and edge computing tracks that build-out, while the cybersecurity desk follows the defensive side of the same story. The semiconductors that run the models remain a separate front in US policy; the European Union this week signed the US-led "Pax Silica" pledge to keep China out of AI-capable chip supply chains, as Euractiv reported.

Business fallout builds

The reviews land as both companies edge toward Wall Street. OpenAI chief executive Sam Altman spoke with Commerce Secretary Howard Lutnick about the Sol release on Wednesday, part of weeks of talks between industry leaders and administration officials. Anthropic joined those conversations too, though its relationship with the White House has been rougher.

The Pentagon labeled Anthropic a national security risk after the company raised ethical and safety objections to AI use in war. Trump ordered federal agencies to drop Claude, and Anthropic answered with a lawsuit that is still moving through the courts. Lutnick wrote to Anthropic on Friday that its work to answer the government's concerns had produced clear progress, and the company said it was "pleased" by the partial return of Mythos.

Trump has also floated a more direct arrangement: the US government taking a stake in leading AI firms, with "pieces" set aside "for the American public, where the American public essentially becomes a partner with the companies." The idea arrived weeks after SpaceX's record June 12 initial public offering, which reset expectations for how the administration views strategic tech listings.

What happens next

OpenAI said it expects Sol to reach more users in the coming weeks, but it stopped short of naming a date. Anthropic said it will keep pressing the government to restore Fable for general use. The 30-day review window in the June order gives a rough clock, though the framework remains a work in progress.

For now, the United States has stumbled into a de facto licensing regime for the most capable models, built through executive action rather than statute. Congress has bills in play, but none has reached the president's desk. Until lawmakers act, the appointees in Washington hold the keys to the next release.

The full Associated Press account of the restrictions is available here.