CAISI to safety-test frontier models

Frontier AI models are getting a new kind of checkpoint. Not a formal licensing regime, and not a law passed by Congress. But something that matters anyway — the U.S. government now has voluntary agreements to test major labs’ most advanced models before those systems go public. On May 5, CAISI, the Commerce Department’s AI testing center inside NIST, said Google DeepMind, Microsoft, and xAI had joined that setup. OpenAI and Anthropic were already in. (nist.gov) ### What is CAISI, exactly? CAISI stands for the Center for AI Standards and Innovation. It sits inside NIST, which is the measurement-and-standards arm of the Commerce Department. Its job here is basically to act as the government’s technical testing shop for advanced commercial AI systems — not to sell policy slogans, but to poke at models and see what they can actually do. (nist.gov) ### What changed this week? The news is the expansion. CAISI announced fresh agreements with Google DeepMind, Microsoft, and xAI on May 5, 2026. Those deals let CAISI run pre-deployment evaluations and related research on frontier models before public release, and they also cover post-deployment assessment. That matters because it turns what used to look like a couple of one-off partnerships into a much broader system. (nist.gov) ### Hadn’t the government already been doing this? Yes — but with fewer labs and less visible structure. CAISI said these new agreements build on earlier 2024 partnerships with OpenAI and Anthropic, and that those older deals were renegotiated under the current administration’s AI priorities. So this is less “brand-new power” and more “the testing net just got wider and more official.” (nist.gov) ### What does “testing” mean here? It means the government gets access early enough to evaluate dangerous capabilities before the public does. CAISI says it runs pre-deployment evaluations, targeted research, and post-release assessments. It also says developers often provide versions with safeguards reduced or removed so evaluators can p(nist.gov)n you are not only testing the consumer-facing wrapper. (nist.gov) ### Why is cybersecurity such a big focus? Because the fastest-moving fear is not just chatbots saying weird things. It is models getting good enough at finding software flaws, automating intrusion work, or helping less-skilled actors do more damage. Politico tied the urgency to cybersecurity concerns around Anthropic’s new Mythos model, (nist.gov)government. (politico.com) ### Is this mandatory regulation? Not yet. The agreements are voluntary. But voluntary does not mean trivial. If every major U.S. frontier lab is feeding top models into the same federal testing channel, that starts to function like a soft pre-clearance norm. The labs keep legal control over release decisions, but surprise launches get harder when Washington (politico.com)er the biggest players. (nist.gov) ### Why does this favor big labs? Because compliance is work. You need secure sharing, internal eval teams, people who can respond to findings, and enough process discipline to coordinate with government testers. Large labs already have more of that machinery. Smaller labs can still move fast, but this kind of regime raises the premium on documentation, red-teaming, and controlled release plans. (nist.gov) ### So what’s the real bottom line? The U.S. still does not have a full legal approval system for frontier AI. But it now has something more concrete than vague safety promises — a standing technical review channel with all five major U.S. frontier labs in orbit. That does not settle the politics. It does change the launch process. (nist.gov)