President Pursues Oversight Of New AI Models

BY AMRITH RAMKUMAR

President Trump signed an executive order Tuesday asking artificial-intelligence companies to give the administration access to powerful models 30 days before public release, just two weeks after shelving a previous version.

The order, which aims to increase the federal government’s oversight of the technology, also asks national-security and cyber officials to work with agency heads and top tech companies to address software vulnerabilities identified by models like Anthropic’s Mythos, the White House said.

The executive order is a slimmed-down version of the one Trump shelved on May 21. That version included the cybersecurity component but would have asked companies to let the government review models for a longer period, up to 90 days.

Trump said at the time that he didn’t want to impede AI’s economic benefits or set the nation back in the tech race with China. It followed a lastminute intervention by venture capitalist and AI adviser David Sacks, who argued that the voluntary model testing could lead to mandatory regulation in the future.

During a meeting Monday evening in the Oval Office that included Treasury Secretary Scott Bessent and Defense Secretary Pete Hegseth, Trump asked to shorten the window for model review, people familiar with the matter said. Sacks called into the meeting and said he was on board with the compromise, they said.

Bessent has helped lead the administration’s response to Mythos, warning that new AI models could wreak havoc on the financial system. He has urged banks to integrate the models into their cybersecurity work.

The executive order, signed without fanfare, is the administration’s latest shift on AI policy, highlighting how Trump is struggling to balance competing factions within the White House: one side seeking more oversight of models, and the other seeking to tear down all barriers to AI deployment. The technology’s implications for the economy and national security also challenged the Biden administration and have divided Congress, which hasn’t passed AI legislation.

National-security and cybersecurity officials, including National Cyber Director Sean Cairncross, have pushed for some AI restrictions to address the technology’s potential harms, which have spurred growing backlash around the country.

The order “keeps America leading in AI while putting these frontier capabilities to work strengthening our cyber defenses,” said Michael Kratsios, director of the White House Office of Science and Technology Policy, which helped lead work on the order.

Sacks said the order “is a game changer because it allows our AI labs to comply with the voluntary framework without delaying new model releases.”

Many AI companies have said they support the approach the White House has taken and had sent executives to Washington for a signing ceremony two weeks ago when Trump pulled the plug. Microsoft President and Vice Chair Brad Smith said Tuesday that the company welcomed the effort.

Critics said the order falls short of the mandatory regulations needed to truly police dangerous AI systems and gives the administration flexibility not to enforce the provisions in a strict way. “This is underwhelming policy that mirrors the Trump Administration’s broader pattern of creating a wild west environment for AI development,” said Rep. Don Beyer (D., Va.), who coleads a group of lawmakers focused on AI.

The administration has come under pressure to regulate AI following the release of powerful models like Mythos and worries about bad actors using AI to create biological weapons.

The order is a significant win for Cairncross and other officials who pushed for more oversight against Sacks and others hoping to accelerate model deployment at nearly all costs, said Dean Ball, a former Trump AI adviser now at the Foundation for American Innovation think tank.

Ball said he thinks the order is a mistake because it could lead to a model licensing system that could hinder innovation, and it wasn’t clear how the government will decide the criteria for tools that must be reviewed before being released.

While many of the measures are described as voluntary, critics said companies would be compelled to comply.

Also on Tuesday, Anthropic said the administration is letting it expand access to Mythos. The White House previously opposed the move because of security concerns.

Anthropic said it is expanding Mythos access to about 150 companies and organizations around the world, up from about 50 that got initial access in April, so that they can use the model to identify software bugs and patch vulnerabilities.

Those with access now include organizations in more than 15 countries and a range of industries that didn’t previously have access such as healthcare, power and water, the company said.

The initial group included tech giants like Amazon.com, Microsoft, Apple and Alphabet--owned Google.

The executive order is a slimmed-down version of the one President Trump shelved on May 21, which would have sought a review period of up to 90 days.

SHAWN THEW/ PRESS POOL