White House AI Framework Is Final, and It Is Secret

The voluntary federal framework for reviewing frontier AI models is finished, and the public is not going to read it. Around a dozen AI companies met White House officials on the morning of 4 August 2026 to close roughly two months of negotiation, and the administration has no current plans to publish the result. If you build products on top of frontier models, the rules now governing your most important supplier are rules you cannot inspect.

What happened on 4 August

The meeting ran about thirty minutes and was staff-level on both sides. Representatives from roughly a dozen AI companies attended, including Anthropic, OpenAI, Google and Meta, with Fortune reporting that Meta, Nvidia and Microsoft were in the room as well. President Trump did not attend, and neither did Chief of Staff Susie Wiles, science and technology adviser Michael Kratsios, or National Cyber Director Sean Cairncross. The AI companies did not send C-suite executives either. This was a closing formality, not a negotiation.

The framework traces back to an executive order signed on 2 June 2026, which called for a voluntary regulatory framework for advanced AI models within 60 days. Semafor reported that the White House confirmed it hit that deadline on 1 August. A White House source told Spectrum News that after the Tuesday meeting the framework is considered final and is being implemented.

The mechanism at its centre is early access. The executive order asked AI companies to grant the federal government access to frontier models 30 days before those models reach other partners, so they can be evaluated first. The order defines frontier models as highly advanced, industry-disrupting models of substantial interest to national security. Participation is voluntary throughout.

The part nobody outside can check

There are currently no plans to make the framework public, at least officially. That is what separates this from an ordinary regulatory story. A voluntary framework readable only by its participants is a set of rules with no external audit, no way for a competitor to know whether it was applied evenly, and no way for a customer downstream to know what was tested.

Chris McGuire, a senior fellow at the Council on Foreign Relations, put it plainly to Fortune: "We can't have secret, voluntary rules to regulate the most important tech in the world."

What was reported as unconfirmed right after the meeting has since been corroborated by multiple outlets, including the Washington Post, Axios and Yahoo News. The framework is administered by CAISI, the Center for AI Standards and Innovation housed inside NIST, and the 30-day review applies only to closed-source frontier models that score at the frontier on cybersecurity and hacking evaluations. Open-weight developers fall outside the review entirely, which is now a confirmed structural asymmetry rather than a rumour. The underlying problem still stands: the framework document itself remains unpublished, so what is public is what leaked rather than the source text.

Why this reaches your stack

Most builders will read "voluntary federal framework" and correctly conclude that nothing changes this week. Nothing does. The reason to pay attention is slower than that.

A pre-release government evaluation window sits inside the release process of the labs whose models you call. If a 30-day access period becomes routine, it becomes a scheduling input on every frontier launch, which affects when capabilities land in your API and how much notice you get. Today that is invisible to you. It will stay invisible, because the process describing it is not published.

The second-order effect is on evaluation. If you are choosing between providers, one of the things you would reasonably want to weigh is what safety and security testing each model has been through. Public frameworks let you ask that question. This one does not. Model choice stays a matter of vendor claims and your own testing, exactly as it was before, which is worth knowing when you read a provider's security page and wonder how much of it is externally verified.

There is also a competitive question inside the now-confirmed CAISI details. Federal review formally applies only to closed models, so the practical difference between open-weight and closed AI models has grown a real regulatory dimension: closed labs face a mandatory pre-release evaluation window that open-weight labs do not. Qwen3.8-27B is a concrete example of an open-weight release that sits outside that review entirely.

What to do with this

Nothing urgent. Three things are worth doing anyway.

Keep your own model evaluation notes, because you cannot outsource the question to a framework you are not allowed to read. If a model is going into a product where mistakes cost you, test it against your own cases rather than relying on a provider summary.

Do not overreact to the open-weight exemption either. It is confirmed now, but a regulatory asymmetry that exists on paper does not automatically show up in a given provider roadmap or pricing this quarter, and rebuilding a supplier strategy around it before an actual launch is affected is how you end up migrating twice.

And keep the general point in view: your exposure to a model provider is a real dependency with real risk surface, and it is a dependency you now understand slightly less well than you did last week. That argues for the boring hygiene of keeping your prompts, your data and your evaluation harness portable, so switching providers stays a decision rather than a project.

State-level rules are moving just as fast as federal ones: Colorado's AI chatbot law is one example of a state imposing its own disclosure requirements.