White House to Extend AI Safety Framework to Open Models

A White House official told WIRED that the administration's new AI safety framework, which right now deals only with the closed models developed by the likes of Anthropic and OpenAI, is expected in the coming months to extend to open models as well. The trigger, per the same official, is capability: once open weights reach the same "frontier" level as Anthropic's Mythos-class models and OpenAI's GPT-5.6, they would be pulled into the same pre-release federal safety testing.

That is a real shift from the picture two weeks ago. When Fortune covered the White House briefing that Meta, Nvidia, Microsoft, OpenAI, Anthropic and a variety of smaller companies attended in early August, the framework was described as excluding open models entirely, and the administration said it did not plan to release the document publicly. The WIRED reporting flips half of that: open models are no longer permanently outside, they are provisionally outside pending capability.

The practical stakes fall on a small list of labs. If the bar is Mythos-class or GPT-5.6-class capability, the developers realistically anywhere near that line for open weights are Meta and the Chinese frontier labs whose releases have kept this policy question live. A voluntary framework that only bites at "frontier" hands Anthropic and OpenAI a de facto regulatory moat their open-weight competitors would have to consciously cross to trigger, and it gives US officials a lever over how far any given open release actually goes. Three tracked experts in our Who's Who directory shared the story.

What WIRED's account does not settle is the mechanics. The framework itself remains unpublished, so nobody outside those White House meetings actually knows what "frontier" means as a testing threshold, what a failed pre-release review triggers, or what a lab that declines the voluntary process is meant to do next. The expansion claim rests on a single unnamed White House official, so the timing ("in the coming months") is directional, not a commitment.

For open-weight labs, the useful move is to engage now, while the definitions are still being written and while "you cannot govern what you cannot see" is a bargaining position. Once the frontier bar is set publicly, arguing with it gets harder.