Google, Microsoft, xAI give early AI access

Frontier AI models are getting a new kind of dress rehearsal — one run by the U.S. government before the public sees them. On May 5, Google DeepMind, Microsoft, and xAI signed agreements letting the Commerce Department’s AI testing arm evaluate some unreleased systems ahead of launch. That sounds bureaucratic, but the stakes are simple: the government wants earlier visibility into whether the most capable models can help with cyberattacks, biological misuse, or other national-security problems. The gap until now was coverage — OpenAI and Anthropic were already in, but several other major labs were not. ### What actually changed? The immediate news is the expansion itself. NIST’s Center for AI Standards and Innovation — CAISI — said the new agreements bring Google DeepMind, Microsoft, and xAI into its pre-deployment evaluation program, where government researchers can test frontier models for national-security risks, not broad product regulation. ### Who was already in? OpenAI and Anthropic were already part of an earlier round of agreements announced in August 2024. Those deals let the U.S. AI Safety Institute work with the companies on evaluations and safety research for advanced systems. So this week’s move does not create the program from scratch — it widens the circle to include more of the companies building the biggest models. ### Why does “before launch” matter? Because timing is the whole point. Once a model is public, the government is reacting to a capability that users, attackers, and competitors can already probe at scale. Pre-deployment testing gives officials a shot at seeing risky capabilities earlier — I mean the testing window moves upstream, closer to the moment when companies still have time to change safeguards. ### What is CAISI, exactly? CAISI is the Commerce Department center inside NIST that serves as the government’s main contact point for testing and collaborative research on commercial AI systems. It is the renamed successor structure around the U.S. AI Safety Institute effort, and it focuses on evaluation first — measuring capabilities, risks, and safeguards — instead of writing product rules line by line. ### Is this mandatory oversight? Not in the classic sense. These are agreements, not a new law forcing every lab to hand over models. The program still runs on cooperation from companies. But the practical shift is real — what started as voluntary safety commitments is hardening into a more regular government testing channel for the biggest U.S.-linked model developers. That gives more leverage than one-off meetings after a launch. ### Why are Microsoft and xAI notable here? Google DeepMind is an obvious frontier-model builder. xAI is too. Microsoft is a little different because it is both a model developer and a platform giant with deep ties across the AI stack. Its inclusion suggests the government is thinking of Microsoft as a frontier AI actor. This last point is an inference from Microsoft’s role in model development and distribution. ### What does this not do? It does not create a formal licensing regime. It does not guarantee the public will see the test results. And it does not settle the bigger political fight over who should regulate advanced AI and how hard. The catch is that early access only matters if the tests are meaningful and the companies act on what they find. ### Bottom line? The U.S. government just got earlier access to more of the country’s biggest AI models. That is not full-blown AI regulation. But it is a clear step away from “trust us” launches and toward standing federal scrutiny before the most powerful systems go live.