US AI Labs Undergo Government Security Review Before Launch

The US Department of Commerce's CAISI has signed access agreements with Google DeepMind, Microsoft, and xAI for pre-launch security and capability evaluations of unreleased models.

• Including Anthropic and OpenAI, which have already been cooperating, all major US frontier AI labs are now entering a voluntary pre-review system.
• CAISI has so far conducted more than 40 model evaluations, and it's reported that some involved red-team testing on versions with lowered safety guardrails.
• New York Times reported that the Trump administration is considering a separate mandatory pre-review executive order.

While voluntary agreements currently form the core of the system, there is speculation that a mandatory framework could be introduced alongside it in the future.