President Trump signed “Promoting Advanced Artificial Intelligence Innovation and Security” on June 2, asking frontier developers to voluntarily hand the federal government up to 30 days of pre-release access to their most capable models. The order explicitly forbids mandatory licensing or preclearance. It also builds every piece of scaffolding such a regime would require.

The earlier May draft asked for 90 days; per CFR, Trump pulled it over competitiveness concerns. What survived is shorter on its face and structurally more aggressive underneath. Within 30 days, the Committee on National Security Systems and the Secretary of War must prioritize cyber defense of federal systems. Within 60 days, Treasury, NSA, and CISA, consulting NIST and the National Cyber Director, must design the voluntary framework itself and a classified benchmarking process for identifying “covered frontier models.”

That term goes undefined. Holland & Knight flags the gap: agencies retain wide discretion over thresholds, benchmarks, and categories, and the NSA Director alone makes the designation call. There are no published criteria for “trusted partners” either, and the vulnerability clearinghouse sits at Treasury rather than CISA, a placement industry counsel at A&O Shearman and Skadden have already noted as contested.

The timing isn’t subtle. The same day Trump signed, Anthropic expanded its cyber-capable Mythos model from roughly 50 organizations to 200. OpenAI’s GPT-5.5-Cyber is already in deployment. CFR reads the order as an attempt to engineer a defender’s window: patch critical infrastructure before adversaries weaponize the same capabilities.

Tracking tools like LemonLime have flagged August 1 as the date worth watching. That’s when the classified benchmarking design lands, and when “voluntary” either holds or quietly stops being the operative word.

Sources

Sources