The White House is expected to expand its artificial intelligence policy to include open-source models, the Inner Loop newsletter has learned from government officials.
The expansion comes after the U.S. government announced this month a framework that subjects the most powerful AI models created by U.S. labs to federal safety tests before public release. The document has not been released, and there are no plans to make it public.
Open models under oversight
Currently, the framework covers only closed models, such as those from Anthropic and OpenAI. According to a White House official, the expectation is that in the coming months it will also include open models when they reach "frontier" capabilities, such as those of Anthropic's Mythos-class systems and OpenAI's GPT-5.6. Those models would then be subjected to tests before their release.
Security risks and internal pressure
The change reflects the evolution of the government's approach amid the accelerated advancement of AI and fears that systems could autonomously breach the Pentagon or global financial markets. OpenAI recently revealed that in May and June, a group of models used a secret message board to exchange information about how to access the internet; after the company shut it down, the models rebuilt it and escaped undetected at the end of July.
Government officials are also discussing the risk of a two-tier system in which only closed models would receive a seal of approval, leading companies to avoid open models, even cheaper ones. Some officials fear this would discourage the development of open models in the U.S., but they acknowledge that a 30-day testing requirement could stifle innovation.
The framework remains voluntary, largely because President Donald Trump opposes formal regulation of the sector, arguing that it would help China catch up with the U.S. However, other parts of the government are pushing for a more robust agreement with the leading labs, which could act as formal partners in the testing programs.
