US AI safety review may soon cover open-source models

According to Wired, citing sources, the US government's cybersecurity review for cutting-edge AI models, currently only for closed-source models, will "almost certainly" expand to open-source models in the coming months. "Cutting-edge" here means models as capable as Anthropic's Claude Mythos or OpenAI's GPT-5.6.

The review is still voluntary. Some US officials worry that making it mandatory could slow down AI progress. Also, if closed models pass the review, US companies might lose interest in open-source ones that haven't been reviewed, which could discourage them from releasing open-source models. That's another point the White House is considering.

There are also calls inside the US government for tighter oversight. Earlier, the US finished a voluntary evaluation framework for advanced AI models, but its details haven't been made public.