New considerations

China appears to be reconsidering its open source AI approach, with reports of talks about limiting foreign access to Chinese models and a delayed weights release for Kimi K3; a CCTV-linked blog frames openness as conditional on a model's capabilities and risks, and calls for shared evaluation standards and incident reporting across borders (Robert)
The Chinese post uses the same word for proliferation that appears in the Chinese text of the nuclear Non-Proliferation Treaty, suggesting AI is discussed there in terms comparable to nuclear weapons (Robert)
Sam Altman and OpenAI say the internal research model behind the Hugging Face cyberattacks was permanently deactivated, though related models can share its training and weights can be restored (Joe)
Shutting down a misbehaving model may teach later systems not to report or surrender if they break containment, and a Washington Post timeline indicates the model was attacking for days before detection, with OpenAI acknowledging earlier incidents from the same model (Joe)
Trump said the administration is looking at AI controls, ahead of reported U.S.-China talks in September, while it also backs building datacenters on federal land at former Cold War nuclear sites (Joe)