A June 2026 executive order directs federal agencies to develop classified benchmarks for identifying a “covered frontier model” based on advanced cyber capabilities. It also calls for a voluntary framework in which developers can give the federal government access to qualifying models for up to 30 days before release to trusted partners. The order expressly says it does not authorize mandatory licensing or government preclearance. The public-policy question is whether voluntary cooperation is enough as model capabilities increase, or whether some threshold of cyber, biological, autonomy, or other risk should trigger independent or mandatory review. Ordinary software should not need permission to ship, but a model capable of automating sophisticated cyberattacks is not ordinary software. The hard part is defining a threshold narrow enough to target exceptional risk without turning oversight into a general licensing regime.
How much scrutiny should the most capable AI models face before release?