White House AI Order Shifts to Voluntary Review After Policy Whiplash

The White House is preparing to release an executive order that would establish a voluntary framework for federal agencies to review cutting-edge artificial intelligence models before they reach the public, according to five people familiar with the initiative who spoke to multiple news outlets.

The directive represents a significant retreat from mandatory pre-deployment testing floated earlier this month, instead asking developers to submit certain advanced models to a coalition of national security and civilian agencies as far as 90 days before public release. The shift follows weeks of conflicting signals that left policy observers uncertain about the administration's direction on AI oversight.

One draft version splits the order into two sections: a cybersecurity component outlining a voluntary clearinghouse formed by the Treasury Department and other agencies to identify and remediate security vulnerabilities in unreleased models, and a "covered frontier models" portion defining which AI systems would qualify for the voluntary review framework. The cybersecurity section also calls for expanded hiring at the US Tech Force, a body of engineers recruited to modernize government computer systems.

Major AI companies including OpenAI and Anthropic have been engaging with the White House on the executive order, according to sources with knowledge of the discussions. The Department of Commerce's National Institute of Standards and Technology announced earlier this month that major tech firms would share unreleased versions of their models with the government for national security and public safety evaluation, though that announcement has since been removed from the Commerce Department's website.

(The policy development comes amid a broader clash between Anthropic and the Pentagon over the company's efforts to limit certain military uses of its technology, as well as the subsequent controlled release of Anthropic's powerful Mythos model to a select group of companies through Project Glasswing.)

The expected executive action maintains what the administration characterizes as a broadly pro-innovation approach while tightening scrutiny over developers of advanced AI systems and addressing risks of catastrophic harm. The voluntary nature of the framework marks a departure from more prescriptive regulatory approaches, reflecting ongoing tension within the administration between national security imperatives and industry competitiveness concerns.

The Trump administration had taken a more hands-off regulatory posture on AI until Anthropic unveiled its Mythos model, which the company says can exploit cybersecurity vulnerabilities at unprecedented speed. That release prompted national security officials to express concern about what one expert called "superhuman hacker" capabilities that could threaten financial systems, power grids, and government infrastructure. Both OpenAI and Anthropic are now granting businesses and governments special early access to their latest models to strengthen cyber defenses.