White House finalizes a closed-door AI evaluation framework

The White House completed its voluntary framework for evaluating advanced AI models by the deadline set in a June 2 executive order, an official told Axios on August 3. "Discussions with industry about next steps are underway," the official said. A day later, Axios reported that the administration does not intend to release the text, citing three people familiar with the discussions. The companies inside the process get the details. Nobody else does.

What is in the framework

The finished document has not been posted anywhere, so what is known about its contents comes from the June order that commissioned it. That order asked for a structure developers can use to ask the government whether a model still in training would be covered. It also called for rules governing government access to models before they are released, spanning confidentiality, cybersecurity, insider risk, intellectual property, and nondisclosure, and for a list of the "trusted partners" who get early access.

Parts of the machinery are formally classified. The benchmarking process for advanced cyber capabilities is, and so is the capability threshold that decides which models are covered at all, to be shared with developers and researchers "as appropriate." The framework itself is not classified. It is simply not being published. A White House official put the distinction plainly to Axios: "Just because things are unclassified that doesn't mean we are going to broadcast them to everyone."

Who was in the room

Anthropic, OpenAI, and Google gave feedback on a draft before the deadline, and the White House official said the administration is engaging "many more" industry partners than those three. Staff-level meetings with companies took place the day after the completion announcement. Open-source models came up in those meetings, though in what way has not been reported, and Nvidia staff took part. Jensen Huang had been in Washington the previous week, meeting Commerce Secretary Howard Lutnick and President Trump.

The White House declined to comment on the August 4 story about keeping the text private. The European Union also declined. The United Kingdom did not respond.

Why a framework now

The administration has been gating frontier model releases one at a time for months. OpenAI shipped GPT-5.6 publicly on July 9 only after a roughly 12-day restricted rollout at the government's request, with the Commerce Department's Center for AI Standards and Innovation running additional tests before the public launch. In late June the administration cleared Anthropic's Claude Mythos 5 for more than 100 trusted US institutions after blocking it over jailbreak concerns. A standing framework is the natural successor to that case-by-case handling, but the reporting does not say whether the completed document formalizes it, and nobody outside the process can compare the two.

Does this affect you

Participation is voluntary, in the administration's telling, and no date has been given for when companies will start operating under the framework. Chatbot subscribers have nothing to change and nothing new to read. The parties who feel this are the labs. A company inside the process has a finished set of rules to work against. Everyone left out, including other companies, policymakers, researchers, and US allies, is left guessing at how the policy will be implemented, which trusted partners get early model access, and whether any foreign government would qualify.