The U.S. government wants a chance to use top AI models before the public does. On June 2, Trump signed an executive order titled Promoting Advanced Artificial Intelligence Innovation and Security, centered on voluntary government access to frontier models for as long as 30 days before release.
The order gives Treasury, War and Homeland Security 60 days to define a framework for what counts as a frontier model. It points to confidential benchmark testing of cyber offense and defense capability, with the NSA director making the final call. CISA must issue mandatory directives, while Treasury is tasked with an AI cybersecurity intelligence exchange for critical infrastructure such as power grids and hospitals.
The key limitation is explicit: the order says it must not be read as creating a mandatory government license, preclearance or permit requirement. In practice, companies may choose whether to share unreleased models with the government. The offer is cooperation and early trusted access, not a launch veto.
That makes the order a compromise. It gives Washington a formal way to approach the most powerful models while preserving Trump’s deregulatory posture. OpenAI, Anthropic and Google now have to decide how much cooperation is acceptable without handing over their strongest unreleased systems too early.
Sources: CocoLoop; The White House; Scientific American