White House Refuses to Show the Rules It Uses to Vet Powerful AI Models

The White House is keeping the rulebook for its AI safety reviews locked in a drawer. The Office of the National Cyber Director (ONCD) handed over some documents in response to a Freedom of Information Act lawsuit, but it's still withholding the core records, citing national security and trade secrets, according to heise.de.

What did leak out is thin. The Information, which reviewed the material obtained by the civil rights group Protect Democracy, reports the framework runs just seven pages — and its full contents remain unreleased. Most of what did surface is email traffic between the ONCD and the Commerce Department and AI labs, showing which organizations got approved as 'Trusted Partners' for early access to frontier models. The complete partner list? Still hidden.

The whole setup rests on Executive Order 14409, which President Donald Trump signed in June 2026. Under it, developers like Anthropic and OpenAI can — if they want to — give the government a look at their frontier models up to 30 days before a planned release. The order stresses this isn't meant to be a mandatory licensing or preapproval regime. In practice, the industry still has no idea which models count as 'covered frontier models,' what tests get run, or which safety criteria apply. Only a handpicked crew of companies has been briefed.

Protect Democracy filed its FOIA request in August 2026 and sued the ONCD and others that September to pry the documents loose. The group's argument: the executive branch gets to decide which AI products ship and which customers get access, all without congressional authorization or criteria the public can see.

The secrecy has a second cost. Outside researchers can't reproduce the benchmarks, smaller providers can't line their safety work up against public criteria, and favored 'Trusted Partners' may walk away with an informational and market edge. Vivian Dong, a director at Legal Advocates for Safe Science and Technology, put it bluntly: 'The federal government is asking the public to trust a framework for reviewing AI models before they are released, but only the government and AI companies know what the framework says. That is unacceptable, and we will continue to fight in court until the public can see it.'

The White House didn't answer a request for comment, according to The Information. So for now, the public gets to trust a framework it isn't allowed to read.