Trump signs executive order: AI companies can "voluntarily" submit their latest models for government review, did Mythos scare the White House?

On June 2, Trump signed an executive order requiring the establishment of a “voluntary framework”: AI companies may submit their models to the federal government for review of advanced network capabilities up to 30 days before frontier models are released; at the same time, the order calls for an “AI Cybersecurity Information Exchange” to be established within 30 days.
(Background: Bitcoin crashes to “$66k”! 1.78 billion in liquidations, with long and short positions wiped out—bloodbath for bulls and bears, 90% hit; the fear index plunges to 11)
(Additional background: NVIDIA is seeking a “presentation designer” with an annual salary of NT$6.5 million! Why, with the same job doing PPT, can’t AI do it?)

Table of Contents

Toggle

• From letting go to soft oversight
• Mythos scared the White House
• The ceiling of voluntary measures

U.S. President Trump signed an executive order titled “Promoting Advanced Artificial Intelligence Innovation and Safety” on June 2, establishing a “voluntary framework”: AI companies can, up to 30 days before frontier models are publicly released, submit their models to the federal government for review of their advanced network capabilities.

The key is “voluntary.” Companies can say no and will not face any punishment; but voluntary does not mean meaningless.

From letting go to soft oversight

In terms of AI policy, the Trump administration has long taken a deregulation-oriented approach. During the period led by former White House AI adviser David Sacks, the government’s tone was to downplay safety concerns and let the technology run on its own. The appearance of this order indicates that this tone has quietly shifted.

The specific mechanism of the order is as follows: AI companies that are willing to participate may submit frontier models to the federal government 30 days before the models are released. By “frontier models,” it simply means the strongest AI systems currently available—for example, models at the GPT-5 and Claude Opus level. The government will assess whether these models possess “advanced network capabilities,” namely whether they could be used to attack critical infrastructure.

The order also requires that within 30 days, the Secretary of the Treasury, together with the National Cyber Director, the NSA, and CISA, establish an “AI Cybersecurity Information Exchange,” working with industry and operators of critical infrastructure to coordinate vulnerability scanning software, verification, and patching.

It is worth noting that the order is explicit, in black and white: this should not be regarded as mandatory authorization, nor as pre-release review clearance. Simply put, the government has no power to approve or reject—it is only a viewing window, and you can only see it when you choose to open it.

Mythos scared the White House

If you want to find the direct catalyst for this order, Anthropic’s “Mythos,” which was released in a limited way in April of this year, is likely a key factor.

Anthropic said that during testing, Mythos flagged “thousands of high-severity vulnerabilities, including those present in every mainstream operating system and web browser.” In other words, the capabilities this model has are sufficient to make people rethink what could happen if a malicious actor got hold of it.

This outcome also seems to have helped thaw the freeze in the relationship between Anthropic and the Trump administration. Previously, relations between the two were quite tense. Anthropic had gotten into legal disputes with the Pentagon over its plan to use AI for autonomous lethal weapons and large-scale surveillance. Mythos effectively provides a shared language: AI cybersecurity risk has become specific enough that it can no longer be ignored.

Brad Carson, chair of the Responsible Innovation Coalition, said, “The White House has officially been won over by Mythos,” and pointed out that this order shows the Trump administration is taking AI vulnerabilities seriously. Brendan Steinhauser, CEO of the Safe AI Alliance, also said he is pleased to see the government confronting the risks of these models.

Worth mentioning is that even advocacy groups that have long opposed limits on state AI regulations have praised this order. The design of the voluntary framework places very tight constraints on industry, so it is not surprising that support is broad.

The ceiling of voluntary measures

The biggest limitation of this order—and also what is inherent in it—is this.

While Carson and Steinhauser praised it, both clearly called on Congress to enact mandatory protection legislation. The meaning behind this is clear: a voluntary framework cannot solve the real problems. Without enforceability, companies that choose not to share will not be recorded for a violation; companies that choose to share will at best receive some confidential protections. An asymmetric incentive structure makes it difficult to ensure full coverage.

Last month, Google, Microsoft, and xAI agreed to accept CAISI’s pre-release review. This was not because of the executive order; rather, it was a business decision made before the order even existed. The order’s main function is closer to institutionalizing this kind of collaborative relationship than creating an entirely new oversight mechanism.

What a voluntary executive order can do is direct attention to the existence of the problem, not solve the problem itself. As for when enforcement will arrive—that is a matter for the legislative body. And the clocks of Congress and of AI labs are clearly not on the same frequency.