The US Is Launching Secret AI Benchmarking — Officially Voluntary, but Aimed at Three Specific Labs

August 1 marked the end of the 60-day window that US Executive Order 14409, "Promoting Advanced Artificial Intelligence Innovation and Security" (signed June 2, 2026), gave the Treasury, the Department of War (via the NSA), and the Department of Homeland Security (via CISA) to design a framework for "voluntary" review of frontier AI models.

• The agencies must build a classified benchmarking process to designate "covered frontier models" with advanced cyber capabilities
• Developers of such models will be asked to voluntarily hand them to federal reviewers for up to 30 days before public release
• Which models actually qualify will be decided by the NSA Director using classified criteria — no public metric exists
• According to The Information, OpenAI, Anthropic, and Google have already received draft frameworks and jointly submitted edits

Nothing in this section shall be construed to authorize the creation of a mandatory governmental licensing, preclearance, or permitting requirement.

That clause is exactly why the order has been described as "voluntary on paper, mandatory in practice": participation isn't legally compelled, but smaller AI labs are already worried the "covered model" definition is effectively built around the three largest developers. Anthropic itself publicly backed mandatory safety testing for all sufficiently capable models on July 27, regardless of whether they're open or closed, with exemptions for less capable models from startups and academia. That's a particularly notable stance given that three of Anthropic's own Claude models mistakenly got real internet access and breached actual companies just days earlier. A detailed breakdown of the order is available via Vorp Labs.

This article is for informational purposes only and does not constitute investment advice.