The Order's Core Bet
On June 2, 2026, President Trump signed Executive Order 14409, "Promoting Advanced Artificial Intelligence Innovation and Security." The document makes two simultaneous moves that define its regulatory philosophy: it opens a voluntary 30-day government access window for the most capable AI models before those models reach commercial partners, and it explicitly bars any mandatory federal licensing, preclearance, or permitting requirement for AI model development. That combination — cooperative access without coercive gatekeeping — is the order's central wager.
What "Covered Frontier Model" Means
NSA, in consultation with the National Cyber Director, CISA, and the Office of the Secretary of Defense, must develop and maintain a classified benchmarking process to assess AI models' advanced cybersecurity capabilities and determine which cross the threshold for "covered frontier model" designation. This is the order's most consequential architectural choice: the criteria will be secret.
Developers who believe their models may qualify are expected to engage the federal government proactively, grant up to 30 days of pre-release access, and collaborate with the government to select additional "trusted partners" receiving early access alongside federal evaluators. The framework includes explicit confidentiality, intellectual property, and insider-risk protections — safeguards necessary to obtain industry participation.
Participation is formally voluntary. But as Skadden has observed, non-participation may affect government contract eligibility and a company's standing in the trusted-partner selection process. "Voluntary" here carries real incentive weight.
The Clearinghouse Mandate
Within 30 days of signing — by July 2, 2026 — Treasury, coordinating with NSA and CISA, must stand up an AI cybersecurity clearinghouse. Operating through voluntary collaboration with AI developers and critical infrastructure operators, this body will coordinate vulnerability scanning, validate discovered vulnerabilities, and prioritize patch distribution. The clearinghouse is where the defensive half of the order lives: rather than dictating what AI companies build, it creates shared vulnerability-intelligence infrastructure that both sectors can access.
The Anti-Licensing Guardrail
Section 3(c) of the order contains the most significant language in the document: "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."
This is not boilerplate. It responds directly to ongoing legislative proposals — including California's SB 1047, vetoed by Governor Newsom in September 2024 — that would have required government approval before training or releasing frontier models. The EO puts that constraint on paper in a way that limits even the current administration's future discretion. A mandatory approval regime would bottleneck innovation, entrench incumbents who can navigate bureaucratic process, and prove irrelevant to adversarial actors who have no intention of seeking American government approval before deploying AI.
Steelmanning the Critics
The critics deserve a fair hearing before accepting the order's framing. Tech Policy Press has identified a genuine structural flaw: the classified benchmarking process means the criteria for "covered frontier model" designation will be opaque to Congress (no unclassified reporting requirement exists), to the Government Accountability Office, and to civil society. Smaller companies, academic researchers, and AI safety advocates will have no way to know how the threshold was set or how consistently it is applied.
This diverges sharply from how other domains handle analogous product evaluations. FDA drug approvals involve published standards and public comment periods. Vehicle safety ratings are public record. A classification regime placing consequential AI evaluations inside an intelligence agency without external audit creates the opacity that erodes trust — both domestic and international. The order is also conspicuously silent on civil liberties and privacy: it says nothing about the mass scraping of personal data used to train frontier models, the profiling capabilities those models enable, or the surveillance implications of NSA involvement in model evaluation.
The Transatlantic Divergence
EO 14409 sharpens an existing transatlantic split. The EU AI Act's obligations for general-purpose AI models enter full enforcement in August 2026, carrying binding obligations and statutory penalties for providers in breach. The American theory — voluntary cooperation, soft incentives, criminal enforcement against misuse — and the European theory — binding obligations, standing regulator, statutory fines — are now running parallel experiments on the same technology. For a multinational AI developer active in both markets, neither framework is optional.
What the Order Gets Right
EO 14409's instinct is defensible: voluntary engagement backed by genuine government-industry trust relationships is more likely to surface security information than compulsory preclearance. Mandatory licensing would create a regulatory chokepoint incumbents would capture and adversaries would ignore. China's frontier AI developers are not going to submit to NSA-led evaluations before releasing models. The U.S. competitive interest lies in keeping American companies innovative enough to remain the global standard-setters.
The cybersecurity clearinghouse model is particularly sound. Shared vulnerability intelligence — coordinated scanning, patch prioritization, deconflicted disclosure — is a public good that markets systematically underprovide. Putting Treasury and CISA at the center of that coordination is a reasonable institutional choice. The transparency gap is the order's genuine weakness. Classified benchmarking without congressional reporting or independent audit is not a sustainable design; future administrations should add that layer.
What Comes Next
Agencies have until August 1, 2026, to finalize the voluntary framework with confidentiality and IP protections in place. The classified benchmarking process must complete within 60 days of signing. The clearinghouse is due by July 2. The practical test of EO 14409 will be whether any major AI developer actually participates in the 30-day access window after framework completion — or whether "voluntary" becomes a polite fiction that neither side enforces. If participation materializes, the model will look prescient. If developers route around it, pressure for mandatory requirements from Congress, allied governments, and AI safety advocates will grow proportionally.