White House withholds all 60 pages on its AI model pre-release review framework

The White House has declined to release any records on its voluntary AI pre-release review framework, The Information reported. A letter dated October 30 says all 30 located records, 60 pages in total, are exempt from disclosure under the Freedom of Information Act.

A letter that closes out the agreed production

The letter comes from the Office of the National Cyber Director and was published by the nonprofit Protect Democracy Project. It accompanies a production in Protect Democracy Project v. Office of the National Cyber Director et al., Civil Action No. 26-3064. The defendants named in the letter are ONCD, the Office of Science and Technology Policy, the Commerce Department’s Bureau of Industry and Security, and the Treasury Department.

The production followed a joint status report filed on September 15, in which the nonprofit narrowed its original August 7 request to four categories of records:

• the text of the voluntary frontier AI disclosure and pre-release review framework, which an administration official reportedly said was complete by August 1;
• records given to AI companies that set out the terms of taking part in a clearinghouse called GOLD EAGLE;
• the documents naming the individuals and entities treated as “trusted partners” under the framework;
• records identifying the legal authority used to set up and run the framework and the clearinghouse.

According to the letter, the agencies searched their electronic systems and found about 60 pages of potentially responsive material. They concluded that nothing could be released, even in part, and say this is their full production under the joint report. A Vaughn index, which lists each withheld record with its date, type and the reason for withholding, is attached.

Four exemptions cover everything

The government relies on four FOIA exemptions. Exemption 1 covers national security information; the letter says disclosure could cause grave damage. Exemption 3 covers cyber threat indicators and defensive measures shared under the Cybersecurity Information Sharing Act of 2015. Exemption 4 covers confidential commercial information, here customer identities that submitters treat as private. Exemption 5 covers internal government communications, where the letter cites presidential communications and deliberative process privileges.

The index entries visible in the letter’s first pages are emails dated July 1 and July 2, 2026 and later. The government describes them as ONCD messages to an AI lab or a model vendor naming trusted partners cleared for pre-release access to a covered model. The government says these lists derive from confidential customer lists and are withheld in full. The letter says non-exempt content is too entangled with protected material to separate out.

What the withholding leaves open

The letter does not say whether the framework text itself is among the withheld records. The index rows we could read describe only emails about trusted-partner access. We did not confirm the full index, so the part of it covering the framework document and the legal-authority records remains unverified here.

Protect Democracy filed suit on September 1 and, according to its case page, sought a preliminary injunction the same day. It says the framework was created without legislation, and that the administration has described it only in general terms and has not named the participating companies or the legal basis. Those are the nonprofit’s claims. The government has not, in the documents we read, addressed them beyond asserting the exemptions. The case page also lists a first, redacted production from OSTP on September 8.

Reporting on the case also described a coalition of more than two dozen organizations from the left and right urging the White House to publish the framework. Separately, one aggregator post citing a reporter said the White House had agreed to release the framework text by October 30. We could not confirm that from the court documents, and the October 30 letter says the opposite outcome.

Rules, safety and ethics: A framework nobody outside can inspect

The framework is voluntary and, by the administration’s own account, covers certain advanced models whose code is not public. For AI companies, the practical question is what they are agreeing to when they take part; for the public, it is who decides which models are released and on what grounds. With the text, the participants and the legal basis all still undisclosed, neither question can be answered from public sources.

The government’s position is that parts of this material are sensitive on national security grounds and parts are confidential business information. Those assertions have not been tested by a court. The Vaughn index exists so the court and the plaintiff can challenge them, and the letter itself reserves the right to add further exemptions. The next step belongs to Protect Democracy and the judge handling the case; the sources we read do not say what either will do, or when.