Why we're watching this: The delayed EO is the clearest signal yet that the US government has no agreed framework for evaluating frontier AI models before release. With Anthropic's Mythos and OpenAI's GPT-5.5 Cyber already capable of exploiting security vulnerabilities at scale, the governance gap is widening in real time.
Key Takeaways
- Trump delayed signing an AI security executive order on May 21, citing language he said “could have been a blocker” for US AI leadership over China
- The EO would have required AI companies to share advanced models with the government for 14 to 90 days before public release for security evaluation
- The unofficial reason for the delay: major AI and tech CEOs couldn’t attend the signing ceremony on short notice, per reporting from The Verge’s Alex Heath
- Anthropic’s Mythos, which can exploit cybersecurity vulnerabilities at unprecedented speed, was a direct catalyst for the EO effort
The Trump White House postponed a planned executive order on AI and cybersecurity on May 21, just hours before the signing ceremony was expected to take place. The order had already been delayed multiple times and would have established a voluntary framework requiring AI companies to share unreleased frontier models with the government for security review.
Trump told reporters he “didn’t like certain aspects” of the order and was concerned the language “could have been a blocker.” He framed the delay in competitive terms: “We’re leading China, we’re leading everybody, and I don’t want to do anything that gets in the way of that lead.”
Behind the scenes, the more immediate reason was logistical. Major AI and tech CEOs had been invited to attend the ceremony but could not travel to Washington on short notice, according to a note seen by Axios and confirmed by The Verge’s Alex Heath on X.
Update: White House is postponing this (lots of CEOs couldn’t make it on such short notice) https://t.co/dObLG5z08V
— Alex Heath (@alexeheath) May 21, 2026
“I really thought that could have been a blocker and I want to make sure that it’s not.” — Donald Trump, US President
The stalled EO came directly out of concerns triggered by Anthropic’s Mythos model, which the company has not released publicly and is instead making available through a tightly controlled consortium called Project Glasswing.
Mythos is capable of finding and exploiting cybersecurity vulnerabilities at a speed that prompted both government and industry to accelerate the policy conversation.
OpenAI is similarly offering early access to its latest models to businesses and governments to help shore up cyber defenses ahead of public release.
One draft version of the EO included two sections: a cybersecurity portion establishing a voluntary clearinghouse between the Treasury Department, other agencies, and AI companies to identify and fix vulnerabilities in unreleased models, and a “covered frontier models” section defining which AI systems would be eligible for the pre-release review framework.
The pre-release window had been a central point of contention, with one version proposing up to 90 days and AI companies pushing for as few as 14.
The Department of Commerce’s NIST had separately announced earlier this month that major tech companies would share unreleased AI models with the government for national security evaluation, but that announcement has since been removed from the Commerce Department’s website.
No formal pre-release review framework for frontier AI models currently exists in the US. The EO delay means Mythos-class models can continue to be deployed to selected partners without any government visibility into their capabilities before broader release.
The Trump administration had maintained a hands-off posture on AI regulation until Mythos changed the calculus. The delay does not close the door on the EO, but it does confirm that agreeing on the pre-release review window, and getting the industry into the room at the same time, remains unsolved, for teams tracking AI governance developments via Relve, an AI trends and tools intelligence platform.
