Trump AI framework covers only closed-source frontier models

The Trump administration's finalized frontier AI framework carves out a narrow universe of models it wants a look at before release, and leaves everything else alone. Axios reports the definition of a 'covered frontier model' is closed-source, with state-of-the-art capabilities and national-security risks. There is no clear definition of what counts as state-of-the-art or what counts as a national-security risk. Open models are excluded from the regime, and the framework explicitly says nothing in it should be interpreted as restricting open models once they have been released.

The mechanism is voluntary. Companies willing to participate hand federal agencies secure access to a covered model for up to 30 days before it goes public. During that window, employee access is restricted, the models sit in high-security environments, and there are detailed logs on who is accessing them. The rules also cover confidentiality, cybersecurity, insider risk, intellectual property and nondisclosure. Federal agencies were tasked with designing the framework by August 1, 2026, and the Axios reporting on the finalized details landed shortly after that deadline.

Why this matters if you are not a frontier lab: the shape of US AI oversight for the next few years just narrowed. A closed-source developer can now build a formal, private relationship with the federal government before shipping, which reads as a trust signal for enterprise and government buyers. An open-weight developer sits outside the regime entirely, with no obligations and no seat at the table. The framework treats those two worlds as different kinds of things, not as points on the same scale.

The honest caveats are heavy. The definitions are soft, and the White House isn't planning to publish the framework, so most of what is known comes from officials briefing reporters. There is no public list of companies that have signed on, no compute or benchmark threshold that says which model is state of the art, and no reporting on what actually happens if the 30-day review turns up a concern, whether the government can slow a launch or only advise. The upside worth watching is who quietly opts in first, because that will tell you more about who the administration counts as a trusted partner than any of the language in the document itself.