Trump Administration Issues Revised AI Executive Order After Abandoned Signing
The Executive Order That Wasn’t Becomes the Executive Order That Could
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.
These were the words of President Donald J. Trump on May 21, 2026 following the decision to abruptly abort the signature of a new executive order on artificial intelligence (AI). For those counting, that was 15 days ago as of this writing. Two days ago as of this writing, the very executive order that got all dressed up with nowhere to go finally had its moment. On June 2, 2026, 13 days after Trump’s quote, the White House posted “Promoting Advanced Artificial Intelligence Innovation and Security” on its website. The relatively short (by typical EO standards) document creates a voluntary program whereby the federal government will gain access to “covered frontier models” 30 days in advance that will be used to scan federal and critical infrastructure systems for cyber vulnerabilities. Stemming from the April 2026 announcement and non-release of Mythos by Anthropic and its assertions that it posed far too great a threat to cybersecurity to warrant release to the public, the new executive order is an interesting move toward federal government involvement in frontier AI model development from an Administration previously allergic to any such intervention.
Executive orders go through multiple iterations before they are released publicly, and this one was likely no exception. After its aborted signing two weeks syne, the original copy was reported to contain a framework for the federal government to “vet the national security risks” of frontier models implying the development of AI testing frameworks, standards, and a very real intervention in the AI industry. The current order contains no language around “vetting” or “evaluating” models but says instead that model providers can voluntarily work with the federal government to determine whether they meet an as-yet undefined threshold of “covered frontier model” and provide the federal government access to such models 30 days before their release. It does not say the government will engage in testing or evaluation of the models or how this will happen.
(i) engage the Federal Government to determine whether model(s) under development meet the designation of “covered frontier model”;
(ii) provide the Federal Government with access to covered frontier models, subject to appropriate confidentiality, cybersecurity, insider-risk, and intellectual-property protection, use, and nondisclosure requirements, for a period of up to 30 days before they plan to release such models to other trusted partners; and
(iii) collaborate with the Federal Government to select trusted partners that will have early access to covered frontier models to promote secure innovation and strengthen the cybersecurity of critical infrastructure.
This publication strives to provide clarity and engaging prose on emerging technology and the policies made around it. On May 26, 2026 (10 days ago as of this writing), we published a piece on the executive order that wasn’t and highlighted some clarity that process provided the AI industry, even in silence. The contention of the piece was that it was clear the Administration’s position on AI will be almost entirely driven by Silicon Valley. Ten days later, the Administration signs the very order it claimed got in the way of America’s lead in AI. What are observers, analysts, investors, and just plain regular people to make of this?
The same Silicon Valley forces that likely caused Trump to abandon the first signing attempt are almost certainly still in place. There are, however, real signs that users and organizations are increasingly uncomfortable with the level of control around AI and this may be the legacy of Mythos.
Commenting on AI policy making has become a treacherous endeavor, but much like King Henry I urge all of you “Once more unto the breach, dear friends, once more.”
Talk and Action
A feature of many Trump Administration AI policies is the conflict between what they say and what they will pay for. The Genesis Mission, announced the week of Thanksgiving 2025, called for lofty national-level goals on AI infrastructure with echoes of the Apollo program. It tasks federal agencies that absorbed massive cuts to their non-defense budgets to fund this generational investment in AI with…a bake sale?
In the same way, the newest of the AI executive order family has some shortfalls that could render it moot absent some real action directed by the White House.
First, the program the EO announces is voluntary, which should surprise no one if you paid attention in civics class. Executive orders are not laws, so they only apply to the federal executive branch. What they say can have repercussions on industry, particularly if you do business with the federal government, but they are not binding and do not come with funding. So, the creation of a program by executive order that envisions industry participation is, by definition, voluntary. It is not without precedent that the tenants of an EO can become law (see the progression of the DHS Post Quantum Cryptography Roadmap to NSM-10 to the Quantum Cybersecurity Preparedness Act). If Congress were to undertake such a legislative action, this idea would then have both teeth and money, making it considerably more effective. As it stands, this is not the case and the likelihood of Congressional action on this topic is near zero.
While the EO uses words that sound very promising like using “covered frontier models” to scan for cyber vulnerabilities across critical systems, it possesses exactly zero ability to compel any AI company to participate in the program. Anthropic itself had the opportunity to release Mythos to the federal government for just such a purpose in February and opted instead to provide it to other billion-dollar companies. A huge danger for this EO is that the AI companies simply choose not to play. If that happens, the EO is dead on arrival. Given the current tension between the Pentagon and Anthropic, it would be understandable if it and other AI companies simply opted out.
Second, the EO refers to “covered frontier models” but does not define what that means. This sounds like wonky policy speak, but it matters. Having a solid definition on what terms mean prevents lawyers from the other side from using a minor semantic difference to argue they are excluded from coverage under the policy or law. Unless a very tight definition of what “covered frontier models” means is delivered soon, legal teams for every AI company will have ample time to change the terminology they use to make their version of a frontier model increasingly difficult to define. This could also render the EO moot because it could mean that the major providers are “exempt” from the process because they do not meet the policy definition of “covered frontier model.”
Third, the EO says the federal government will work with the large AI companies to use their frontier models to ensure national security, which is noble and hard to argue with. However, the original reporting from 15 days ago said the EO would establish a process for the federal government to “vet the national security risks” of frontier models. There is a significant delta between the government vetting models and the government working with AI companies for 30 days to scan for cyber vulnerabilities. This is almost certainly the change that occurred over a 15-day period. Vetting models would require the development of non-existent model testing methods and likely a lengthy standardization process on the part of NIST. It would have also been politically unpopular because it is hard to argue that this is not direct government intervention in technology development.
When the history is written, there is a great chance this AI EO results in very little that wasn’t already planned. That this arrangement between the AI companies and the federal government was already underway before this EO is a real possibility, which would mean the Administration is taking a victory lap on something that was already in the works. Not without precedent in presidential administrations, but hardly the kind of impact that the AI world was looking for.
Another Brick in the Wall
The Trump Administration is weaving a real tapestry of AI policy actions in its still young term. This EO seems to be a way to save face after aborting the signature of a different version that reportedly included a program to vet models before their release. It’s also an answer to the Mythos non-release that has made many people nervous about AI development. It’s hard to argue with collaboration on better cybersecurity, but it is also hard to miss what this order does not do. But here’s what is clear:
There is real appetite for some level of governance around AI model use and development, and the White House is getting closer to that than it has ever been.
Cybersecurity and AI are closely tied together in a way they weren’t before thanks to Mythos.
Federal policy on AI continues to be heavily influenced by Silicon Valley.
The delta between talk and action in AI policy making remains.
This EO sets the precedent for more government involvement in frontier model development.
What’s also clear is that there’s a real chance this EO is just for show. It looks and sounds good, but there are several potential failure nodes here that could render it moot. Similar to the Genesis Mission, words will get us neither to AI safety nor AI innovation. The president talked about leading China and not wanting to do anything to mess up that lead. I agree. But that lead is not assured, and it will require actions, not words. Action means getting serious, putting funding behind smart federal programs, creating vetting and testing tools, and defining what national security risks from models look like. So far, we’ve not even defined what we mean by a covered frontier model.