In late May, tech CEOs received an invitation to the White House. There was going to be a signing ceremony for a landmark AI executive order, and senior officials held a press briefing the morning of the planned announcement to walk through it. Then, hours before the cameras rolled, President Trump pulled the order off the table. He told reporters it could hurt American companies racing against China. "We're leading China, we're leading everybody, and I didn't want to do anything to get in the way of that lead," he said.
On Tuesday, June 2, he signed it anyway. Quietly, in private, with no CEOs in the room.
The order that emerged, titled "Promoting Advanced Artificial Intelligence Innovation and Security," asks the companies building the most capable AI systems in the world to give the federal government a look at those systems up to 30 days before they release them to anyone else. It sets up a classified process for the National Security Agency to decide which models are powerful enough to qualify. And it builds an entire cybersecurity apparatus around the premise that frontier AI is now a national security asset and a national security threat at the same time.
What it does not do is force a single company to participate. That distinction is the whole story.
The Government Wants a Head Start on the Most Powerful Models
The centerpiece of the order is in Section 3, under the heading "Secure Frontier Model Deployment." It directs the Treasury, the NSA, and the Cybersecurity and Infrastructure Security Agency (CISA) to build two things within 60 days.
The first is a classified benchmarking process to measure the advanced cyber capabilities of AI models and set a threshold above which a model becomes a "covered frontier model." The NSA director makes that designation. The criteria stay secret.
The second is a voluntary framework through which a developer can do three things: ask the government whether a model in development meets the covered-frontier threshold, give the government access to that model for up to 30 days before releasing it to other trusted partners, and work with the government to choose which outside partners get early access at all.
In plain terms, the government is asking for a private preview window. Before a lab ships its next flagship model to enterprise customers, Washington wants first look, framed around one question: can this thing be used to break into critical systems, and if so, who else should see it before the public does.
The trigger for this is not hypothetical. In April, Anthropic's Mythos Preview model demonstrated a superhuman ability to find critical vulnerabilities in some of the world's most-used operating systems, a capability that sent what NBC News described as shockwaves through Washington. An AI that can autonomously discover zero-days is a defensive gift and an offensive weapon depending on who holds it. The order is the administration's attempt to get on the right side of that timing.
The Word That Changed Everything Was "Voluntary"
The version of this order that almost got signed in May reportedly asked for a longer pre-release review window. According to reporting from Axios, the administration's AI and crypto czar, David Sacks, pushed to cut that window down to 30 days and to make the entire mechanism voluntary rather than mandatory. The final text reflects exactly that.
Section 3 closes with language that reads less like policy and more like a promise to industry:
"Nothing in this section shall be construed to authorize the creation of a mandatory governmental licensing, preclearance, or permitting requirement for the development, publication, release, or distribution of new AI models, including frontier models."
The testing relies on what the order repeatedly calls "voluntary collaboration" from companies like Anthropic, OpenAI, and Google. The government can ask. It cannot compel. As NBC News put it, the order makes early access "a request, not a rule."
That framing is a direct answer to the fear Trump voiced when he killed the May draft. A mandatory pre-clearance regime would have meant American labs waiting on a federal sign-off while Chinese labs shipped freely. By stripping out the mandate, the administration gets the optics of oversight without the friction of regulation. Whether labs actually hand over unreleased models when there is no penalty for declining is the open question the order does not answer.
The Cybersecurity Machine Underneath
Most of the coverage has focused on the frontier-model access. The larger share of the order is about hardening government and infrastructure against AI-powered attacks, with aggressive deadlines attached.
Within 30 days of the order, the following are due:
- The Committee on National Security Systems must prioritize cyber defense of classified national security systems.
- The Secretary of War must do the same for Department of War information systems.
- CISA must issue Binding Operational Directives to speed up cyber defense of civilian federal systems and expand AI-enabled defensive tools, including making "covered frontier models" available, where appropriate, to agencies, state and local authorities, and operators of critical infrastructure such as rural hospitals, community banks, and local utilities.
- The Treasury, working with the National Cyber Director, the NSA, and CISA, must stand up an AI cybersecurity clearinghouse that coordinates vulnerability scanning, validates the vulnerabilities it finds, and prioritizes the distribution of patches.
Section 4 turns to enforcement. It directs the Attorney General to prioritize prosecution under existing federal computer-crime statutes against anyone who uses AI to break into a system without authorization, including anyone who deploys AI agents to unlawfully access data that then gets used for a crime. That clause lands squarely on the kind of agentic attacks the security community has watched escalate all year, from poisoned packages to autonomous exploit generation.
There is even a small tell in the final section about how this White House sees the stakes: the order specifies that the cost of publishing it "shall be borne by the Department of War."
The Other Side Sees a Toothless Compromise
Not everyone reads a voluntary framework as a win.
Critics who wanted real pre-deployment testing argue that "voluntary" is another word for "optional," and that a lab under competitive pressure to ship has every incentive to skip a 30-day government review it can legally ignore. A safety mechanism that depends entirely on goodwill, in this view, is a press release dressed as policy.
There is also the access problem from the other direction. Handing the federal government an unreleased frontier model, even for 30 days, means exposing a lab's most valuable intellectual property to people outside the company. The order tries to address this with what it calls "appropriate confidentiality, cybersecurity, insider-risk, and intellectual-property protection," but labs have spent the year watching the White House treat Anthropic's Mythos as a supply-chain risk while simultaneously blocking the company from sharing it. Trust runs in both directions, and it is thin.
And the provision that lets the government help "select trusted partners" for early access raises a different concern. A federal hand in deciding which companies see the most powerful models first is, functionally, a federal hand on the competitive scales. The order presents this as a security measure. Rivals may experience it as industrial policy.
What This Means for Practitioners
For engineers and researchers, the immediate effect is close to zero. Nothing in the order changes how you train, fine-tune, or deploy a model today. There is no new compliance burden, no registration, no filing.
The medium-term effect is worth watching. If the major labs do cooperate, model release calendars could shift. A 30-day government window, layered on top of red-teaming and staged rollouts, could mean the gap between a model finishing training and reaching your API key gets a little longer for the most capable systems. The classified benchmarking threshold also matters: where the NSA draws the line for a "covered frontier model" will determine whether this touches only a handful of flagship releases or a wider band of open-weight and frontier models.
This is the third major move in a policy arc the administration has been building all year, after the national framework calling for federal preemption of state AI laws and the order it scrapped at the last minute in May. The direction is consistent: light-touch on innovation, heavy-touch on security, and a clear preference for partnership with industry over rules imposed on it.
The Bottom Line
Trump signed an order asking the most powerful AI companies in the world to show the government their best models before anyone else sees them, then wrote into the same order that those companies are free to say no. It is oversight built on a handshake.
The bet is that the labs will cooperate because cooperation buys them goodwill, government contracts, and a seat at the table when the rules eventually do harden. The risk is that the first time a lab is racing a Chinese competitor to ship, the 30-day courtesy gets quietly skipped, and an order with no enforcement mechanism discovers it has nothing to enforce.
Treasury Secretary Scott Bessent and Vice President JD Vance have both grown more invested in governing how AI spreads. They now have a framework that asks for everything and requires nothing. Whether that is shrewd or hollow depends on a single variable the order leaves entirely to the companies: whether, when asked, they actually say yes.
Sources
- Promoting Advanced Artificial Intelligence Innovation and Security (Executive Order, full text) (Jun 2, 2026)
- Trump signs order seeking early access to powerful AI models before release, NBC News, Jared Perlo (Jun 2, 2026)
- Trump signs AI executive order asking companies to give government early access to models, CNBC (Jun 2, 2026)
- Trump dodges AI rules for now with latest executive order, Axios (Jun 2, 2026)
- Trump Signs an Executive Order to Vet Top AI Models for National Security Risks, U.S. News / AP (Jun 2, 2026)
- Trump signs AI executive order to give government early look at new models, CBS News (Jun 2, 2026)