The White House walked roughly a dozen artificial intelligence companies through its finalized frontier model review framework in closed staff-level meetings Tuesday, then declined to publish the document. OpenAI, Anthropic, Google, Meta, Nvidia, Microsoft and a handful of smaller developers were in the room; Fortune first reported Microsoft’s attendance, and CNBC confirmed the session through a White House official speaking on the condition of anonymity, who said the administration had been coordinating with a broader group of industry partners.

The framework is the operational output of Executive Order 14409, which President Trump signed June 2 and which set an August 1 deadline for two parallel tracks: a classified benchmarking process led by the National Security Agency with Treasury and the Cybersecurity and Infrastructure Security Agency, and a voluntary pre-release engagement process for what the text calls “covered frontier models.” According to multiple sources briefed on Tuesday’s meeting and cited by Axios, that designation applies only to closed-source models with state-of-the-art capabilities and national security risks. Neither term is defined. Open models are excluded, and the text says nothing in it should restrict them after release.

Companies that opt in submit a model up to 30 days before public release. During the window, employee access is restricted, the weights sit in high-security environments, and detailed access logs are maintained. Firms were encouraged Tuesday to submit models close to release rather than early-stage ones.

What’s classified is the interesting part: the capability thresholds that trigger review, and the mechanism by which the NSA Director makes the determination.

The executive order also explicitly forecloses “the creation of a mandatory governmental licensing, preclearance, or permitting requirement.” So this is voluntary, unpublished, and administered by the intelligence community against criteria no outside party can see.

Chris McGuire, a senior fellow for China and emerging technologies at the Council on Foreign Relations, called the decision to withhold the document “baffling.” He added that “we can’t have secret, voluntary rules to regulate the most important tech in the world.” Analysts at Tech Policy Press pressed the same point more procedurally, arguing that “the categories of capability that trigger designation, the evidentiary standard applied, and the process by which the NSA Director makes the determination should all be stated in unclassified form.”

The gap won’t stay empty for long. The Great American Artificial Intelligence Act of 2026, introduced earlier this year, is already being floated as a vehicle for transparency requirements and independent audits. The historical rhyme is the 2018-era Section 230 debates, which began as voluntary industry accommodations and ended as a decade-long congressional preoccupation once the private arrangements proved illegible to everyone outside them.

Sources