US government gains pre-release access to AI models from Google DeepMind, Microsoft, and xAI for national security testing

The Center for AI Standards and Innovation (CAISI), part of the US Department of Commerce, signed new agreements with Google DeepMind, Microsoft, and xAI to test advanced AI models for national security risks before public release. These agreements build on earlier deals with Anthropic and OpenAI.

CAISI has already run more than 40 evaluations, some on unreleased models. Testing occurs in classified environments, and AI labs provide model versions with reduced safety guardrails to support the assessments.

The expansion reflects growing capability of AI models at finding and exploiting security vulnerabilities, according to the agency's stated rationale.