Trump's 30-Day Voluntary Model Review Goes Active Under the New AI Security Order

The voluntary review window at the center of President Trump's June 2 executive order is no longer a paper exercise. Three weeks after the order took effect, OpenAI confirmed it is working with the Office of the National Cyber Director and the White House Office of Science and Technology Policy to test GPT-5.5 and a specialized GPT-5.5-Cyber variant under the new framework — the first concrete evidence that "Promoting Advanced Artificial Intelligence Innovation and Security" has moved from signing ceremony to active implementation.

The order asks frontier developers to voluntarily hand the federal government early access to their most capable models for up to 30 days before broader release, so national-security and cybersecurity reviewers can probe them. It does not require it. And that distinction — between an invitation and a mandate — is the whole story.

What the order actually does

The mechanics are now coming into focus. Within 60 days of signing, the Treasury Department, the National Security Agency, the Cybersecurity and Infrastructure Security Agency, and the National Institute of Standards and Technology are directed to build a classified benchmarking process to assess a model's "advanced cyber capabilities" and decide whether it counts as a "covered frontier model" subject to the voluntary window. On a separate 30-day clock, CISA must issue Binding Operational Directives to harden civilian federal systems, and the Office of Management and Budget must determine whether any existing grant funding can flow to developers building AI vulnerability-detection tools.

But the heart of the order is deliberately optional. As the text states, nothing in it authorizes "a mandatory governmental licensing, permitting or preclearance requirement" for developing, publishing, releasing, or distributing an AI model. Participating developers "may" provide access; the government reviews and advises; deployment decisions stay with the company.

The administration cites Anthropic's recent Fable 5 shutdown as the kind of outcome the framework is meant to catch — a deployment pulled after Amazon researchers breached the model's guardrails — though notably that episode was resolved without any federal mandate at all.

The 90-to-30 cut

The order Trump signed is meaningfully weaker than the one his own administration drafted. The original version set the review window at 90 days, and Trump was reportedly prepared to sign it at a May 21 ceremony. That signing was scrapped overnight after Elon Musk, Meta CEO Mark Zuckerberg, and venture capitalist David Sacks called the president directly to block it, arguing the longer window would slow American AI relative to China.

The window was cut to 30 days. The result is a partial pivot. Trump's earlier AI order had revoked President Biden's AI safety directive, stripping out mandatory pre-deployment testing. The June 2 order quietly concedes that frontier AI carries genuine national-security risk — while pointedly declining to restore the binding guardrails it had removed.

Industry welcomed it. OpenAI, Anthropic, and Google all publicly endorsed the order, with OpenAI CEO Sam Altman saying it struck an appropriate balance and Google's Kent Walker calling it an important step. That near-unanimous praise is itself a tell about how much the framework asks of them.

What critics see missing

Civil-society analysts are far less impressed, and their objections cluster around what the voluntary design leaves out. Writing for Tech Policy Press, critics argued the order "disregards transparency, due process and democratic accountability." Because the benchmarking criteria will be classified, the public "may receive little information about why one model is designated for testing while another is not."

The accountability gaps are specific. There is no requirement for unclassified summary reporting to Congress, no defined role for the Government Accountability Office, and no Freedom of Information Act pathway to the benchmarks or their results. The Council on Foreign Relations and other reviewers note the order also narrows its lens almost entirely to cyber risk, sidestepping privacy, civil liberties, surveillance, and data-brokerage harms.

The deeper critique is structural: a regime built on voluntary participation has no enforcement floor. A company that declines to submit a model faces no penalty. A company that submits one and ignores the government's advice faces no penalty either. The Fable 5 case, held up as a success, was a self-administered shutdown — which underscores rather than resolves the question of what happens when a developer disagrees with reviewers.

The federalism fault line

The voluntary federal approach now sits awkwardly atop a patchwork of binding state law. California's Transparency in Frontier AI Act and New York's Responsible AI Safety and Education Act both impose mandatory disclosure and safety-framework requirements on frontier developers that go further than anything in the June 2 order. Congress has twice rejected attempts to attach a federal preemption moratorium that would have blocked such state rules, and no comprehensive federal AI statute exists.

That leaves an inversion of the usual pattern: the states are the ones with mandates, while Washington offers a 30-day invitation. For developers, the practical compliance burden may end up being driven less by the executive order than by Sacramento and Albany — and the unresolved preemption fight is the most likely place the next AI-policy battle gets fought.

What to watch

Three markers will reveal whether this framework has teeth. First, the classified benchmark due within 60 days — roughly the first week of August — will define which models even qualify, and how transparent its criteria are. Second, the rate of voluntary participation: OpenAI is in, but whether Meta and xAI, whose leaders lobbied the window down, actually submit models will test the order's reach. Third, the preemption fight in Congress and the courts, where the gap between voluntary federal review and mandatory state law could force a reckoning the executive order was designed to avoid.

For now, the order is live, the first model is in review, and the central tension is unresolved: an administration acknowledging frontier-AI risk while declining to require anyone to do anything about it.