The days of AI giants releasing powerful models with a “trust us, it’s safe” approach are coming to an end. Google, Microsoft, and Elon Musk’s xAI have officially joined a partnership that gives the U.S. government early access to their unreleased AI models. These models will go through pre-deployment evaluations designed to catch potential national security risks before they go public.
Google, Microsoft and xAI Open US government AI models safety reviews
The Center for AI Standards and Innovation (CAISI), part of the Commerce Department, is the group holding the magnifying glass. They’ve already run about 40 reviews on cutting-edge models, some of which haven’t even been seen by the public yet. While OpenAI and Anthropic were already on board, this new expansion brings the rest of the industry’s heavy hitters into the fold.
Microsoft has even confirmed they are working with government scientists to probe “unexpected behaviors” by using shared datasets and stress-testing workflows. It’s a big change toward transparency for companies that usually keep their “frontier” tech under lock and key.
Why the sudden pressure? Meet “Mythos”
The urgency behind these agreements isn’t random. A new model from Anthropic called “Mythos” has caused a minor panic in Washington and corporate America. Reportedly far ahead of current tech in terms of cybersecurity, Mythos has officials worried it could supercharge hackers. Anthropic is so concerned themselves that they are restricting access and haven’t felt comfortable releasing it to the public yet.
What happens behind closed doors?
When CAISI gets their hands on a model, they don’t just use the version you’d see on your phone. Developers often provide versions with the safety guardrails stripped away so the government can truly see what the model is capable of when pushed to its limits. From probing for cyberattack potential to assessing military misuse, the goal is to identify threats before a massive rollout
The post Google, Microsoft & xAI to Give US Government Early Access to New AI Models: Testing Safety Before Public Launch appeared first on Android Headlines.
