Washington sets up a preliminary review of the most powerful AI models

Resulting from a June decree, the system provides for government access of about thirty days before the release of models deemed critical to national security.

A decree signed on June 2, 2026 mandated the U.S. administration to build a procedure for identifying and examining so-called "covered border" models, the most capable systems, in terms of cyber and national security risks. Key deliverables were expected by August 1st: a classified measurement process to determine which models cross the threshold, and a window of approximately 30 days of government access before publication.

The Treasury, the War Department via the NSA and the Department of Homeland Security via CISA are involved in the development of these cyber capability assessment frameworks. The participation of developers is presented as voluntary, with guarantees of confidentiality and intellectual property protection — several observers nevertheless describe it as voluntary on paper and mandatory in practice.

The main friction point is the opacity of the device: since the capacity thresholds are classified, companies cannot know exactly what triggers an examination. However, the vast majority of mainstream tools remain below the target line, which only concerns the top of the market.