White House Finishes Its Frontier AI Review, Keeps the Details Quiet

The White House has finished the framework that will govern how the federal government inspects the most capable AI models for cybersecurity risk, meeting a deadline of August 1. It has not said what the framework contains. A staff-level meeting brought the major labs together to walk through the finished document, and the substance has stayed behind closed doors.

The effort traces to an executive order President Trump signed on June 2, which set up a voluntary channel for what the government calls covered frontier models. Under it, developers can hand the government advance access to a model before release, for a window of up to 30 days, so officials can probe its cyber capabilities before the public gets its hands on it. OpenAI, Anthropic and Google are the labs expected to take part first.

Who does the testing

The reviewing work falls to the Center for AI Standards and Innovation, a body inside the National Institute of Standards and Technology, working alongside the National Security Agency. CAISI already describes itself as the government's main point of contact for testing commercial AI systems, and it has run evaluations with several labs over the past two years. The new framework formalizes that arrangement for models at the frontier.

The word doing the heavy lifting is voluntary. Nothing here compels a company to submit a model, and nothing published so far spells out what a review looks for, what would count as a failure, or what happens to a model that raises alarms. Officials confirmed the framework is done and met its deadline. They stopped there.

Trust without a paper trail

That silence is the story. A review process the public cannot read asks for a good deal of faith, both in the government running it and in the companies choosing whether to opt in. It arrives as Congress considers far blunter instruments, including a bipartisan bill that would hand Washington an emergency shutoff for frontier systems, and after Sam Altman's visit to Washington as the outline of this regime was taking shape.

A pre-release look at a powerful model is a reasonable idea on its face. The government has a real interest in knowing what a system can do before millions of people can prompt it. But a framework nobody outside the room can examine is hard to judge, and harder to trust. For now the public is being asked to accept that the review exists and works, without being shown either.