Trump signs scaled-back AI order giving government 30-day voluntary preview of frontier models

What the order does

President Donald Trump signed an executive order on Tuesday creating a voluntary framework for the federal government to vet powerful new AI models before they are released. The order tasks the Office of the National Cyber Director with developing a process to share information about software vulnerabilities identified by AI systems with operators of critical infrastructure, including banks, local utilities and hospitals. Under the new rules, tech companies would be asked to share their AI models with the government for a review up to 30 days before a public release, a window narrowed from the 90 days in an earlier draft that some industry officials had pushed to shorten to as little as 14 days.

I didn't like certain aspects of the original order.

The order stops short of imposing mandatory review requirements. Participation by AI companies would be voluntary, though leading firms are expected to take part. National security agencies will be required to bolster cybersecurity abilities and create a "cybersecurity clearinghouse." Within 60 days, the Treasury Department, the National Security Agency, the Cybersecurity and Infrastructure Security Agency, the National Institute of Standards and Technology and White House officials must develop a classified benchmarking process to assess advanced cyber capabilities of AI models and decide when a model should be treated as a "covered frontier model."

The road to Tuesday's signing

The executive order had been months in the making. Trump had been scheduled to sign a 90-day version of the order on May 21, but abruptly rejected that draft just hours before the planned White House signing ceremony. According to Axios, the White House postponed the ceremony following pressure from tech industry insiders. The president later told reporters he had misgivings about the original text. On Monday, Trump took part in a small, high-level White House meeting where he and his advisors agreed on the new scaled-back order, which was signed in a private ceremony on Tuesday.

Internal administration battle

The order emerged from a three-way internal fight that has paralysed federal AI policy. Three factions have been vying for control: the Commerce Department, which has been quietly building civilian testing partnerships with AI companies; national security officials who want intelligence agencies to evaluate frontier models before release; and pro-industry aides who argue that any regulation risks slowing American AI leadership. The conflict has been described by people involved as a "knife fight." Former White House AI czar and current adviser David Sacks and National Economic Council deputy director Ryan Baasch pushed for language prohibiting the creation of mandatory government licensing. Sacks has continued to play an influential role from his new perch outside the White House, and the abrupt cancellation of the earlier executive order occurred after his involvement.

The Mythos catalyst

The internal fight has been intensified by Anthropic's model called Mythos, which has been shown to be effective at finding security weaknesses in computer code and hacking into networks. The model discovered more than 10,000 zero-day vulnerabilities across every major operating system and web browser. Anthropic declined to release the model to the public, instead inviting a small group of partners to test its capabilities and fix their systems. On Tuesday, the company said it had invited 150 more organizations to join the program, which it calls Glasswing.

I think the idea of testing, particularly for critical infrastructure providers, to be able to identify vulnerabilities and patch them before the capabilities become widely available, there's a lot of sense to that.

Samir Jain, vice-president of policy at the Center for Democracy and Technology, called the order "opaque" at the time, noting it does not give the public much visibility into the benchmarking process. He warned that an opaque procedure allows for the possibility that any administration could exercise arbitrary power over whether, when and how models are released, using security as a pretense to block or handicap a model for political or ideological reasons.

Industry and government partnerships

The order comes after the Trump administration struck a deal last month with Microsoft, Google DeepMind and xAI to review early models of their new AI models before they are released. The Center for AI Standards and Innovation, part of the US Department of Commerce, already has similar deals with OpenAI and Anthropic. On May 5, CAISI announced pre-deployment testing agreements with Google DeepMind, Microsoft, and Elon Musk's xAI, but days later the announcement was removed from NIST's website with no explanation. CAISI staff were told to take the page down but not told why. The federal government recently removed details of that agreement from its website, although it is unclear why.

Advanced AI capabilities make our Nation stronger, but also introduce new national security considerations that require coordinated action across executive departments and agencies.

The executive order is another sign that Trump is moving away from his initial low-regulation approach to AI. One of his first actions as president was to revoke a Biden-era executive order that established standards for safely developing AI. The new guardrails come amid rising fears that the latest AI models can be dangerous, especially in the wrong hands.