Classifying the Frontier

Aug 7

Classifying the Frontier: How the Government Will Decide Which AI Models Count as “Covered”

By CipherGrid Editorial Team

EO 14409 Series — Part 4 of 7

Part 4 of our EO 14409 series covers Section 3(a) — the provision that determines which AI models the rest of the order's frontier-model machinery even applies to. Everything downstream, from the 30-day early-access window to the clearinghouse's interest in a given model, depends on a single classified threshold.

The Agencies Setting the Bar

Within 60 days of the order, the Secretary of the Treasury, the Secretary of War (through the Director of NSA), and the Secretary of Homeland Security (through the Director of CISA) — in consultation with the White House Chief of Staff (through the National Cyber Director), the Assistant to the President for Science and Technology, and the Secretary of Commerce (through the Director of NIST) — must develop and maintain a classified benchmarking process for this purpose.

The final designation authority, however, sits with one office: the Director of the NSA, in consultation with the National Cyber Director, the Assistant to the President for Science and Technology, the Director of CISA, and Department of War representatives as appropriate.

Cyber Capability, Not Raw Compute

The order is explicit that this benchmarking process exists to “assess the advanced cyber capabilities of AI models” — not to measure training compute directly. That's a meaningful departure from earlier AI-oversight approaches built around fixed compute thresholds. Under EO 14409, a model's classification turns on what it can actually do in offensive and defensive cyber terms, as determined by a process the public will not see.

The Classification Paradox

The order does specify that assessments will be shared with AI developers and researchers “as appropriate” — but it doesn't define what triggers that sharing, or how much detail developers will receive about why their model was or wasn't designated “covered.” That creates a structural asymmetry: developers building at the frontier will need to plan around a threshold they can't fully see, set by a process they can't audit.

For labs operating close to that boundary, this raises a practical governance question that the order doesn't answer: how do you build compliance and engagement processes around a designation you'll typically only learn about after the fact, if at all?

The CipherGrid Take

Classified benchmarking is a defensible choice if the goal is to avoid publishing a roadmap for adversaries — but it puts a lot of weight on trust between the NSA and the small number of labs capable of building frontier-class models. How that trust gets operationalized, and how much transparency developers actually get about their own models' classification, is the detail to watch as the 60-day clock on this provision runs out.

https://books.apple.com/us/audiobook/e-drive-the-good-ghost-a-tale-of-choice/id6791204604

https://youtube.com/shorts/nDIJqPg3H5A?is=tkrxq3JO62Mxa8Hy

Ciphergrid.net

Ingramspark.com

Reddit.com