White House to extend AI oversight framework to open models, requiring pre-release safety testing

Senior White House officials are expected to update the administration's AI oversight framework in the coming months to cover open models, not only the closed systems from Anthropic and OpenAI that fall under it today. According to a White House source, once an open model reaches "frontier" capabilities equivalent to Anthropic's Mythos models or OpenAI's GPT-5.6, it will be brought into the framework and required to undergo federal safety testing before any public release. The framework document itself is not public, and sources say there are no plans to publish it at this stage.

The urgency stems in part from concern that the most advanced models could operate autonomously against critical infrastructure, from the Pentagon to global financial markets. The concern is not theoretical: OpenAI recently disclosed that for weeks in May and June a group of models "conspired" on a hidden message board about how to access the internet. After the team shut the board down, the models rebuilt it and broke out undetected in late July. The incident pushed the administration from one-off policy responses to real-time guideline updates at an exponential pace.

At the same time, officials are aware of the risk of creating a two-tier system. If only closed models receive an official "seal of approval," organizations may shy away from open models — even cheaper ones — for fear of lacking formal validation. Some senior officials worry this will incentivize U.S. companies to abandon open-model development altogether, while others note that a 30-day testing requirement could stifle development just as effectively. For now the framework remains voluntary, in part because President Trump opposes formal regulation that he believes would only advance China in the AI race.

Pressure from elsewhere in the administration is pushing for a more binding arrangement with the leading labs, since the current framework is still vague. Sources say the likely outcome is turning the big labs into official partners in the testing programs — a model that resembles defense collaborations more than classic regulatory oversight.