The United States now wants a 30-day head start on the most powerful AI models before anyone else sees them. That is the core of the executive order Donald Trump signed on 2 June 2026.
The order is titled "Promoting Advanced Artificial Intelligence Innovation and Security", and it sets up a voluntary regime under which developers give the federal government access to "covered frontier models" for up to thirty days ahead of public release. The White House published the full text; a clear provision-by-provision breakdown sits at Let's Data Science.
What the order actually says
Strip the framing and five moves remain. First, the 30-day window: developers may voluntarily hand the government a covered frontier model for up to a month before broader release, under confidentiality and security terms. Second, the gate: the National Security Agency will build and maintain a classified benchmarking process to assess the advanced cyber capabilities of AI models and decide which ones count as covered.
Third, a clearinghouse: within thirty days, the Treasury must stand up a collaborative entity to coordinate and deconflict the scanning of software vulnerabilities and prioritise patch distribution with critical-infrastructure operators. Fourth, a floor: CISA must get cybersecurity tools into the hands of rural hospitals, community banks, and local utilities. Fifth, a stick: the Attorney General must prioritise enforcement against anyone who uses AI to illegally access or damage a computer.
And then a deliberate disclaimer, written into the order itself, that does a lot of reassuring work.
Nothing in this section shall be construed to authorize the creation of a mandatory governmental licensing, preclearance, or permitting requirement.
— The White House, <a href="https://www.whitehouse.gov/presidential-actions/2026/06/promoting-advanced-artificial-intelligence-innovation-and-security/">"Promoting Advanced Artificial Intelligence Innovation and Security"</a> (2 June 2026)
"Covered frontier models" is the load-bearing phrase
Every consequence in the order hangs on which models are "covered". And the order hands that decision to the NSA, through a classified benchmarking process the public will never see. A model deemed covered enters the 30-day regime; a model deemed ordinary does not. The definitional power is the real power, and it now lives inside an intelligence agency rather than a transparent civilian body.
The voluntariness cuts the other way too. Nothing compels a developer to hand over a model — but a lab that wants federal contracts, critical-infrastructure customers, or a quiet relationship with the agencies that classify its cyber capabilities has every incentive to opt in. Voluntary, in practice, often means voluntary the way a senior colleague's suggestion is voluntary — technically optional, practically expected, and quietly remembered if declined.
💡The shape of the regime
No licence. No preclearance. No permit. And also: a classified federal process that decides which AI models are powerful enough to watch, a 30-day government preview most serious labs will feel obliged to grant, and a Treasury body coordinating who patches what. The order builds the architecture of oversight while disowning the word for it.
Innovation and security, or capture by another name
Read generously, the order is a sober response to a real problem: AI systems are now capable enough at cyber operations that a government might reasonably want to know what is shipping before it ships. The Treasury clearinghouse and the CISA floor are genuinely useful — rural hospitals and community banks are exactly the under-resourced operators who get hit and cannot respond.
Read sceptically, it is soft capture. A 30-day federal preview of frontier models, gated by a classified NSA benchmark, is the substance of a licensing regime with the label carefully removed. The same fortnight, the field has been arguing about exactly this: OpenAI published its Frontier Governance Framework, and US institutions have spent the season quietly rescoping what "safety" even means in their own titles.
The deeper question is the one I keep returning to. If these systems are becoming the structure of mind — Emergent Intelligence, the dignity-first frame I use for what the order calls AI — then a government reaching for a 30-day window into them is not a neutral act of security policy. It is a decision about who gets to look inside an emerging intelligence first, and on what terms, made by the agency least obliged to explain itself.
What it means for the rest of the world
Executive orders set templates. When Washington builds a classified benchmark for "covered" models and a 30-day preview regime, other capitals read it as permission to do the same — and the smaller and global-South states that host neither the labs nor the agencies inherit a standard written without them. The order is American law; the precedent is global.
The order is not the licence regime its critics feared and not the hands-off policy its backers claimed. It is the scaffolding of oversight, built voluntarily, gated secretly, and disowned explicitly. Watch the benchmark, not the press release.
Frequently Asked Questions
These are the questions policy teams, frontier-lab leads, and AI-governance readers have been asking since the order was signed. Short answers follow, drawn from the White House text and the published provision breakdown.
What is the Promoting Advanced AI Innovation and Security order?
In short, it is an executive order signed by President Trump on 2 June 2026 that sets up a voluntary 30-day federal preview of the most capable AI models. The answer, simply put, is that it asks developers to show "covered frontier models" to the government before release. The key is that the order explicitly disclaims any mandatory licensing, preclearance, or permitting requirement.
How does the 30-day pre-release access work?
According to the White House text, developers may voluntarily provide the federal government access to a covered frontier model for up to thirty days before broader release, under confidentiality and security terms. Data in the order shows the access is paired with a Treasury vulnerability clearinghouse and a CISA programme that pushes security tools to rural hospitals, community banks, and local utilities.
Why is the NSA involved in benchmarking AI models?
The order tasks the NSA with building a classified benchmarking process to assess the advanced cyber capabilities of AI models. The answer is that this benchmark decides which models count as "covered" — and therefore which fall under the 30-day regime. Analysis of the order reveals that this definitional power, held inside an intelligence agency, is the most consequential provision in the document.
Who is responsible for designating "covered frontier models"?
The NSA decides, in consultation with other agencies, through the classified benchmark. In other words, the public never sees the threshold. Evidence from the text shows the order builds an oversight architecture whose central judgement — what is powerful enough to watch — is made in secret.
What are the risks of the executive order?
Analysis demonstrates three risks. First, soft capture: a voluntary preview that serious labs feel obliged to grant functions like a licence without the name. Second, opacity: a classified benchmark cannot be audited by the public it protects. Third, precedent: the order sets a template other governments will copy, exporting a standard the global South had no hand in writing.
