White House Proposes Voluntary Pre‑Release Review for Frontier AI

The White House has briefed leading AI developers on a draft executive order that would create a voluntary framework for government review of so‑called "covered frontier models" before they are released to the public. Reporting on the briefings surfaced in multiple outlets on May 20–21, 2026.

One draft under discussion would ask labs to notify and give agencies an early look — in one version up to 90 days before public launch. Industry sources have pushed back, seeking much shorter windows, with some companies reportedly offering roughly two weeks instead of 90 days.

The Office of the National Cyber Director (ONCD) hosted briefings that included AI firms such as OpenAI and Anthropic, along with cloud providers, semiconductor and cybersecurity firms, and some banks, according to people briefed on the sessions. Those meetings aimed to explain the draft’s contours and gather industry feedback.

The draft names several agencies that would play roles in the review process — the ONCD, National Security Agency, Office of Science and Technology Policy, Cybersecurity and Infrastructure Security Agency and others — with different offices responsible for defining which systems qualify as "frontier" and for conducting classified capability assessments.

Although described as voluntary, the proposal would bring government security and cybersecurity testing closer to the model‑release cycle. Administration officials and some national security advisers see that as a way to detect dangerous capabilities and vulnerabilities before models are exposed to actors who might exploit them.

For AI labs and cloud partners the move raises operational and commercial questions. Even a voluntary pre‑release window could force changes to internal release timelines, contractual terms with cloud hosts, and handling of sensitive training or model details that companies regard as trade secrets.

The policy shift follows concern inside government about models that can surface cybersecurity flaws and other risks. Anthropic’s decision earlier this spring to withhold and tightly control a highly capable model was cited by officials as a catalyst for revisiting how the government and industry manage high‑risk releases.

Not everyone in Washington supports a long review delay. Critics argue a 90‑day window could slow U.S. innovation and leave American firms at a disadvantage to foreign competitors. President Trump, speaking as the administration delayed a planned signing, said he did not want a measure that would be "a blocker" to U.S. competitiveness with China.

The definition of a "covered frontier model" is a central unsettled point. Drafts reportedly leave threshold setting to a mix of security agencies and classified processes, a structure that raises questions about transparency, appeal rights for developers, and how to protect proprietary information during technical review.

Cloud providers and partners are a second layer of stakeholders. Some versions of the draft would permit or require vetted critical‑infrastructure operators and cloud hosts to test or access models ahead of release so they can prepare defenses and compatibility measures, a step that could alter commercial deployment playbooks.

Policy experts say the voluntary frame could be a bridge between the hands‑off approach of earlier months and a future, legally binding regime. If the draft hardens into policy or statute, the government’s role could shift from advisory testing to mandatory gating, depending on subsequent White House choices or congressional action.

For labs and cloud partners the immediate task is practical: work through how much pre‑release access they can safely give, what technical artifacts they will share, and how to preserve trade secrets and competitive agility. The next public steps are likely further consultations and possible revision of the draft before any signing.