Google, Microsoft and xAI agree to give U.S. government pre-release access to frontier AI models for review

Google, Microsoft, and xAI just signed on to let the U.S. government inspect some of their most advanced AI models before the public sees them. That sounds bureaucratic, but the real shift is simple — frontier AI releases are starting to look less like surprise product launches and more like controlled security events. The government wants earlier visibility into what these systems can do, especially around hacking, military misuse, and other national-security risks. On May 5, the Commerce Department’s Center for AI Standards and Innovation, or CAISI, made that arrangement official. ### What exactly did they agree to? CAISI, which sits inside NIST at the Commerce Department, said Google DeepMind, Microsoft, and xAI will provide access for pre-deployment evaluations and targeted research on frontier models. The deals also cover post-deployment assessment work. In plain English, the government gets a look at important systems before release and can keep studying them after launch. ### Is this mandatory regulation? Not yet. The agreements are voluntary. But “voluntary” undersells what is happening. Once every major U.S. frontier lab is inside the same review channel, that channel starts to feel like a de facto checkpoint for serious model launches. CAISI said these new deals build on earlier arrangements with OpenAI and Anthropic, which were first set up in 2024 and have now been renegotiated under the Trump administration’s AI priorities. ### Why now? Because Washington is getting more nervous about what the newest models can enable. The immediate backdrop is concern over cyber capabilities — especially the possibility that powerful models could help attackers find vulnerabilities, automate intrusion work, or scale up offensive operations. Reuters tied this week’s move to alarm around Anthropic’s newly unveiled Mythos model and its hacking potential. That does not mean Mythos caused the whole policy shift by itself, but it clearly sharpened the urgency. ### What does “pre-release access” really mean? It does not just mean government staff get a demo. CAISI said developers often provide versions of models with safeguards reduced or removed so evaluators can test the raw capabilities and risks. The agreements also support testing in classified environments and let evaluators from across government participate through CAISI’s TRAINS task force, an interagency group focused on AI national-security concerns. Basically, the government wants to know what the model can do when the training wheels come off. ### How big is this testing effort already? Bigger than it may sound. CAISI said it has already completed more than 40 evaluations, including on state-of-the-art models that were still unreleased at the time of testing. So this is not a brand-new review machine being sketched on a whiteboard. It is an existing system that just got broader coverage across the top tier of U.S. AI labs. ### Does this slow product launches? Probably, at least at the margin. If you are a lab shipping a frontier model, you now have stronger incentives to stage rollout plans, prepare red-team materials, document safeguards, and coordinate with outside evaluators before launch. The catch is that none of this looks like a formal licensing regime yet. It is closer to a standing expectation that the biggest labs will give the government a preview window. That can still shape release calendars in a real way. ### Is a tougher system coming next? Maybe. Politico said the White House is considering an executive order that would create a more formal government review process for new AI models. If that happens, this week’s agreements may look like the bridge between loose voluntary safety pledges and a more structured federal vetting system. ### Bottom line? The important change is not just that three more companies signed paperwork. It is that all five major U.S. frontier AI labs now appear to be operating inside the same federal testing orbit. That makes pre-release model review feel less like an experiment and more like the new normal for top-end AI.