Skip to content

Trump Spent Weeks Drafting an AI Order. Thursday, He Scrapped It at the Last Minute.

DS
LDS Team
Let's Data Science
9 min
The order would have asked AI labs to hand frontier models to the government 90 days before public release, a response to Anthropic's Mythos discovering thousands of cyber vulnerabilities on its own. Trump pulled it Thursday afternoon, telling reporters he did not want anything slowing America's lead over China. The plan was already far weaker than what security hawks had demanded.

The CEOs were reportedly told to come to Washington. The order had been in the works for weeks. By Thursday afternoon, May 21, the only thing President Trump had to do was sign it.

He didn't.

Standing in the Oval Office during an unrelated event, Trump told reporters he had pulled the executive order at the last minute. "We're leading China, we're leading everybody, and I don't want to do anything that's going to get in the way of that lead," he said of America's AI industry. "I really thought it could have been a blocker."

With that, the most concrete attempt yet to put federal eyes on frontier AI models collapsed before it began. For the engineers and researchers who build and ship these systems, the near-miss is a preview of the oversight regime that is coming, and a measure of how unwilling Washington still is to actually impose it.

The Order Was Built Around One Model's Wake-Up Call

The draft, reported by NBC News, CNN, and Axios from people familiar with it, had two halves: cybersecurity and frontier models.

The frontier-model section is the one that would have touched daily practice. It would have charged agencies including the Cybersecurity and Infrastructure Security Agency, the White House Office of the National Cyber Director, and the Office of Science and Technology Policy with deciding which systems count as a "covered frontier model." It would then have created a framework for the government to access and evaluate those models before public release.

The mechanism was a voluntary one. Labs would be asked, not required, to share a covered model with the government as much as 90 days before shipping it, and to give early access to certain critical infrastructure providers such as banks. The cybersecurity half aimed at hardening the Pentagon and national security systems, expanding cyber hiring, and pushing AI tools into soft targets like utility companies and rural hospitals.

The trigger for all of it has a name. Less than two months ago, Anthropic's Mythos Preview model demonstrated it could autonomously discover thousands of severe and critical software vulnerabilities, including flaws in leading operating systems and web browsers. Anthropic never released it publicly, sharing access only with a vetted set of companies and agencies through a program called Project Glasswing. We covered that rollout in Anthropic's zero-day hunting AI. The capability that impressed defenders also terrified the government, because the same skill that finds vulnerabilities can be used to weaponize them.

The Government Has Been Circling This for Months

Thursday's reversal did not come out of nowhere. The administration has been inching toward frontier-model oversight, then flinching, for most of the year.

January 2025
Trump repeals Biden's AI order on day one
The scrapped Biden framework had required leading labs to share internal testing results. Trump's approach favored a light touch from the start.
Early April 2026
Anthropic's Mythos Preview spooks Washington
The model autonomously finds thousands of severe vulnerabilities, including in major operating systems and browsers, softening the administration's full-speed-ahead stance.
Early May 2026
A CAISI testing deal appears, then vanishes
A NIST announcement about expanded pre-deployment testing with Microsoft, Google, and xAI is posted, then quietly pulled from the website days later.
May 20, 2026
Vance signals the order is imminent
The vice president says the administration is prioritizing protecting people's data and privacy after Mythos, and is working collaboratively with the tech companies.
May 21, 2026
Trump scraps the signing
He pulls the order at the last minute, saying he did not want to risk slowing the U.S. lead over China.

The CAISI episode is the tell. The Center for AI Standards and Innovation, housed inside the Commerce Department's National Institute of Standards and Technology, had already lined up pre-deployment testing with three major labs. Then the announcement disappeared from NIST's site. We wrote about that arrangement in the CAISI testing deal with Microsoft, Google, and xAI. The pattern is consistent: the administration keeps building oversight machinery and keeps getting cold feet about turning it on.

The Order Was Already a Compromise, and It Still Died

Here is the part worth sitting with. The plan Trump killed was not aggressive regulation. It was close to the gentlest version imaginable.

Vice President JD Vance laid out the rationale a day earlier, on Wednesday. The administration was prioritizing "protecting people's data" and "people's privacy" after Mythos, he said, warning that "a bad actor could use Mythos to target various cybersecurity vulnerabilities." His framing was cooperative, not coercive: "Right now, we're working in a collaborative way with the technology companies, and we're just trying to make sure that the American people are as safe as possible."

Compare that to what the order actually contained:

  • Participation in model testing would have been voluntary, not mandatory.
  • There was no requirement that any model pass a government review before launch.
  • Unlike Biden's repealed order, it would not have forced labs to hand over internal testing results.

In other words, the document asked the most powerful AI companies to please share their most dangerous models early, if they felt like it. Even that was apparently too much friction for an administration that has made beating China its organizing principle.

The Counterargument Has Real Teeth

Trump's reasoning is not empty. American labs and officials have repeatedly pointed to the EU's AI Act, with its stricter requirements, as a drag on innovation, and there is a coherent case that any review step that adds 90 days to a release cycle hands an advantage to competitors who face no such step. If the genuine fear is a Chinese lab shipping faster, a voluntary U.S. framework that only American companies honor could widen exactly the gap it was meant to close.

The counterpressure is just as real. Anti-AI sentiment is rising in polling, and security hawks inside and outside government wanted far more than a voluntary ask, especially after Mythos and OpenAI's GPT-5.5-Cyber showed models finding and exploiting flaws at speeds never seen before. A White House official, asked about the order before it was pulled, would only say that "discussion about potential executive orders is speculation," which is how administrations talk when they have not decided what they believe.

The result is a policy caught between two fears: moving too slow against China, and moving too slow against the models themselves. On Thursday, the China fear won.

The Bottom Line

For people who build AI systems, the takeaway is not that oversight is dead. It is that the federal government has now drafted, staged, and abandoned a frontier-model review process in the span of a single afternoon, which tells you the regime is coming but the timing is hostage to politics. The agencies, the definitions of a "covered model," and the 90-day idea are all written down somewhere in a White House drawer. They will come back, probably after the next model does something frightening in public.

Until then, the only oversight that exists is the kind labs volunteer for. Anthropic's caution with Mythos and the on-again, off-again CAISI deals are the actual policy, and they depend entirely on companies choosing restraint. That worked this week because the most dangerous models are still mostly unreleased. It is a thin thread to hang national cybersecurity on.

Trump said he didn't want a blocker. The question he left unanswered is what happens the first time a frontier model becomes the threat the order was written to contain, and there is no order to point to.

Sources

Practice interview problems based on real data

1,500+ SQL & Python problems across 15 industry datasets — the exact type of data you work with.

Try 250 free problems
Free Career Roadmaps8 PATHS

Step-by-step roadmaps from zero to job-ready — curated courses, salary data, and the exact learning order that gets you hired.

Explore all career paths