The White House is withholding its AI model testing framework, and a court fight is on

Somewhere in the federal government there is a document, reportedly seven pages long, that decides how the most powerful AI models in the country get tested before they reach the public. It has been final since August 2026. Nobody outside government and a handful of AI companies has read it, and as of early November a lawsuit to force its release has not succeeded.

What exists

Start with what is on the record. A June 2026 presidential order, Executive Order 14409, lets AI developers such as Anthropic and OpenAI choose to give the federal government access to frontier models up to 30 days before planned release, per heise online. The order states it does not create a mandatory licensing or preapproval requirement. CBS News reported that President Trump halted an initial signing because he did not want the federal government to slow innovation.

The framework that implements the order is the part that is hidden. Protect Democracy, a nonprofit, says the administration described it publicly only in general terms: it covers certain AI models that do not release their code, and it was finalized on August 1. The contents, the participating companies and the legal basis have not been disclosed.

CBS reported in September that the framework is confidential, and that neither the administration nor the companies are required to release review results or confirm they take part. It added a data point of its own: Sam Altman told Axios that OpenAI submitted its new Astra model for review and called the process "productive."

The lawsuit

Protect Democracy Project filed a FOIA request in August and sued on September 1, 2026, naming the Office of the National Cyber Director, the Department of Commerce, the Department of the Treasury and the White House Office of Science and Technology Policy. Its filings that day included a motion for a preliminary injunction and declarations from several people, including Senator Josh Becker.

The docket moved quickly at first. On September 8, the Office of Science and Technology Policy made a first production of records in redacted form. A joint status report followed on September 15, per Protect Democracy's case page.

Then came the October 30 letter. As reported by The Information and summarized by heise, the National Cyber Director's office declined to release the full draft of the framework, citing national security and protection of confidential commercial information. We could not open The Information's article directly, so we rely on heise's account of it and on reporting that the letter exists.

What was released and what was not

Heise reports the released material is mostly emails from the National Cyber Director's office and the Commerce Department to AI labs, showing which organizations were approved as "Trusted Partners." Still withheld: the full draft framework, the accompanying documents and a complete list of approved companies.

That leaves basic questions unanswered. Which models count as covered frontier models? Which tests run? Which safety criteria apply? The government has briefed only a select group of companies, per heise.

Why the secrecy matters

Vivian Dong of Legal Advocates for Safe Science and Technology told heise that "only the government and AI companies know what the framework says," and added: "That is unacceptable, and we will continue to fight in court until the public can see it."

There are practical reasons beyond principle. Outside researchers cannot reproduce tests whose design is unknown. Smaller developers cannot align their own safety work with criteria nobody has published. And companies that are Trusted Partners may gain an information and market advantage over those that are not, a concern Protect Democracy raises.

There is also a basic accountability problem. Tim Harper of the Center for Democracy and Technology told CBS: "The public has a right to know what the government's interpretation of the law is on these issues." A voluntary program with no legislation behind it is only as strong as the public's ability to see whether it is being followed.

How this fits the wider picture

Other countries are involved too. We covered how the White House restored UK AI Security Institute access to pre-release models after a 17-day freeze. For a running record of which labs have had which models tested, see our pre-release model testing tracker. A framework whose participants are unlisted makes that tracker harder to keep honest.

Congress is the other route. Our tracker of AI safety bills in Congress lists the proposals in play.

Pressure for more visibility is coming from inside the industry as well. CBS reported that in July 2026 more than 1,000 AI company employees asked for a U.S.-backed international effort to develop tools to "deliberately pace the frontier of automated AI development." That request was about international coordination, not this framework, but it shows how far the workforce building these models has moved toward wanting rules that can be seen.

Our take

Voluntary pre-release testing is a reasonable idea, and the fact that labs are submitting models is better than the alternative. The problem is that the program cannot be audited. A framework that is secret, voluntary, built (Protect Democracy argues) without legislation and run with no public criteria cannot be verified as doing anything, and the labs' own statements that the process was productive are the only evidence we have.

What would help is cheap. Publish the seven pages, or a redacted version that states the tests and thresholds, and name the participating companies. If the exemptions are as strong as claimed, a court will say so. We will track the Protect Democracy docket and update this piece when the next filing lands. Two unverified specifics circulating in some coverage, a compute threshold and a reporting deadline, have no source we could open, and we are not repeating them.