Google, Microsoft give US access

Frontier AI is starting to look a little more like aviation or pharma — not because the models are slow-moving, but because governments want to inspect them before they ship. That is the real news here. On May 5, Google DeepMind, Microsoft, and xAI agreed to give the U.S. government early access to unreleased models so officials can probe national-security and public-safety risks before deployment. The gap this tries to close is simple: companies can test their own systems, but governments have threat expertise companies do not. (nist.gov) ### What actually changed? The new agreements sit with the Center for AI Standards and Innovation, or CAISI, inside NIST at the Commerce Department. CAISI said it will run pre-deployment evaluations and targeted research on frontier AI capabilities, with a specific focus on security. (nist.gov)ul models before release. (nist.gov) ### Why does “early access” matter? Because a model can look fine in a product demo and still do dangerous things at scale. The hard problems are not only bias or hallucinations. They include cyber misuse, help with biological or chemical threats, model autonomy, evasion of safeguard(nist.gov)aunch. You need the model, the safeguards around it, and enough time to stress both. (nist.gov) ### Who inside government does the testing? A big piece is TRAINS — short for Testing Risks of AI for National Security. NIST said that taskforce now includes participants from more than 10 U.S. agencies. The point is to combine technical model evaluation with actual government exper(nist.gov)ncies bring the threat models. (nist.gov) ### Is this brand new? Yes and no. The U.S. AI Safety Institute signed earlier agreements in August 2024 with Anthropic and OpenAI for safety research and model evaluations. But this week’s move expands that framework to Google DeepMind, Microsoft, and xAI, and it lands after the(nist.gov) widening pipeline. (nist.gov) ### Why is Microsoft also talking to the UK? Because this is becoming an international testing network, not just a Washington project. Microsoft announced parallel agreements on May 5 with CAISI in the U.S. and the UK’s AI Security Institute. The UK side said the partnership will inclu(nist.gov)cause the same labs ship globally, and governments do not want safety work trapped inside one country’s bureaucracy. (blogs.microsoft.com) ### Does this mean mandatory approval before release? Not yet — at least not from the documents that are public. The agreements describe collaboration, testing, and pre-deployment evaluation, but not a blanket government veto ov(blogs.microsoft.com)stems with national-security implications. That is a meaningful shift from “trust our internal red team.” (nist.gov) ### What does this mean for buyers of AI? It raises the bar on auditability. If the biggest model developers are giving governments pre-release access for risk testing, enterprise customers will start asking smaller vendors tougher questions too — who tested this, against what threat(nist.gov)ooling, where “we evaluated it internally” is starting to sound thin. This last step is partly inference, but it follows directly from the way frontier-model governance is hardening. (nist.gov) ### Bottom line The deal is not that Washington now controls Google, Microsoft, or xAI models. It is that pre-release government testing is becoming normal for the most powerful systems. Once that norm sticks at the frontier, it tends to spread outward — into procurement, regulation, and customer expectations. (nist.gov)