Trump privately signed the order earlier Tuesday, two White House officials familiar with the matter told POLITICO. The president had participated in a small, high-level meeting at the White House on Monday about next steps for the order, according to the two officials and another person familiar with the matter, all of whom were granted anonymity to discuss the closed-door discussions.

The final text asks some AI companies to submit their powerful new models to a voluntary government review 30 days before releasing the products to the public, a pause that would give federal agencies some time to gauge what threats the products may pose to sensitive financial, national security and other computer systems.

An earlier draft of the order had called for a voluntary review as much as 90 days in advance, a provision that some AI industry officials had called too onerous, POLITICO reported last month.

“Advanced AI capabilities make our Nation stronger, but also introduce new national security considerations,” Trump wrote in the order, adding: “We will continue to lead an America First cybersecurity effort that enhances both our national security and our global AI dominance.”

The president had been scheduled to sign the 90-day version of the order on May 21, with a throng of top executives invited to attend. But he abruptly rejected that draft just hours before the Oval Office ceremony, saying he feared it would “get in the way of” U.S. competition with China for AI supremacy.

The earlier draft had received signoff by White House officials at the highest levels and had been reviewed by the tech giants OpenAI, Anthropic and Google. But former AI czar David Sacks warned Trump hours before the planned signing that the order would slow innovation, POLITICO reported at the time.

Sacks was dialed in from the Oval Office for Monday’s meeting, one of the people familiar with the discussion said.

The reversals on the order are just the newest in a series of shifts in the United States’ AI policy, driven by sharply conflicting factions within Trump’s administration, amid rising worries about the technology’s rapidly advancing capabilities.

As with the previous draft, the final order also directs the Treasury Department to coordinate with AI industry officials to establish a “cybersecurity clearinghouse” within 30 days. That initiative will involve a voluntary partnership with AI developers and owners of critical infrastructure to identify and patch cybersecurity vulnerabilities uncovered by the new AI models.

The action also creates a classified benchmarking process to assess the national security implications of advanced AI models and determine which are covered by the policy. This will be overseen by the National Security Agency director in consultation with cybersecurity and technology leaders from the White House, the Cybersecurity and Infrastructure Security Agency and the Pentagon.

Overall, the document is similar to the draft executive order that Trump had planned to sign in May. It even includes identical language denying that the administration was seeking to create a mandatory vetting process: “Nothing in this section shall be construed to authorize the creation of a mandatory governmental licensing, preclearance, or permitting requirement for the development, publication, release, or distribution of new AI models, including frontier models.”

Former Trump AI adviser Dean Ball expressed surprise that so much of the earlier draft remained intact — and said he feared it set the stage for more onerous federal oversight.

“Wow. This EO is almost exactly similar to the leaked text from the EO POTUS chose not to sign because it was too regulatory,” Ball wrote on X. He added that he saw the benefits of the voluntary reviews as “barely articulable” — “what, exactly, is the intelligence community going to do in 30 days to make the models safer?”

The White House’s reversals have added to anxiety among people inside and outside the government that the U.S. is responding too slowly to the dangers posed by AI’s quick evolution. POLITICO reported last month that JPMorgan Chase CEO Jamie Dimon has conveyed concerns to Treasury Secretary Scott Bessent about the speed of the government’s response and the need for coordination with the private sector.

Trump, who got huge support from Silicon Valley’s titans for his second presidential election victory, came into office seeking to ease regulatory burdens on AI in hopes of outcompeting China. But that policy has morphed this year as powerful new models such as Anthropic’s Mythos threaten to undermine cybersecurity in even the most sensitive computer systems.

The administration even floated the idea of imposing mandatory federal reviews of cutting-edge AI models, a process that National Economic Council Director Kevin Hassett had told Fox Business in May would be akin to the Food and Drug Administration’s reviews of pharmaceuticals. But the White House quickly backpedaled on that idea.

In addition to its other provisions, the final order tells the Pentagon to secure its networks within 30 days. It also directs the Justice Department to pursue criminal cases against any individuals who use AI models to hack into computer systems.

The administration will also have 30 days to issue directives requiring that federal agencies step up the defense of federal networks in light of newly uncovered vulnerabilities, and to make AI models with hacking capabilities available to some state and local governments and to critical infrastructure operators.

Details of the policy directive come after weeks of discussions between the White House and industry over how to curtail potentially systemic threats from the technology — especially the risk that AI could help U.S. adversaries find and exploit security flaws faster than anyone can patch them. Those talks followed the release of powerful models such as Mythos, which researchers say has already discovered long-buried vulnerabilities in widely used computer systems.

Mythos’ capabilities have especially alarmed top administration officials since Anthropic announced it in April, and prompted federal agencies to seek access to the model to secure government systems.

Brendan Bordelon, Dana Nickel and Maggie Miller contributed to this report.