Google, Microsoft and xAI Open Their Models to US Government Testing

Three of the largest American AI developers have agreed to let the federal government test their models before public release. Google DeepMind, Microsoft and Elon Musk's xAI signed separate agreements this week with the Center for AI Standards and Innovation, the unit inside the US Department of Commerce that handles voluntary safety review of commercial AI systems.

The Center, known as CAISI, will examine new models for cybersecurity weaknesses, biosecurity risk and the kind of dual-use chemistry knowledge that worries national security officials. Each agreement covers testing, joint research and the development of best practices, and CAISI says it has now run more than forty model evaluations across the industry.

What the agreements cover

The deals build on similar partnerships CAISI struck with Anthropic and OpenAI in 2024. According to Al Jazeera, the new round adds Google DeepMind, Microsoft and xAI to that list, meaning the largest American frontier labs are now all routing pre-release evaluations through the same federal body.

The work goes beyond launch-day checks. CAISI will continue evaluating models after deployment, watching for failure modes that surface only at scale. CNN reports that the evaluations cover "demonstrable risks" with a heavy emphasis on cyber capabilities, the area that has drawn the most attention from policymakers in recent weeks.

Why this is happening now

The agreements arrive during a noticeable shift in tone from the Trump administration on AI oversight. The President came into office promising a deregulatory line, and one of his first acts was to revoke a Biden-era executive order on AI risk. The current discussion of pre-release vetting is, by any measure, a reversal.

The trigger appears to be Anthropic's Mythos model, which is reported to find and exploit network vulnerabilities at a level that rattled officials briefed on it. National Economic Council Director Kevin Hassett told reporters this week that the White House is drafting a separate executive order to formalise pre-release vetting, and compared the proposed regime to FDA drug approval. Bloomberg first reported the FDA framing.

What is and isn't binding

The CAISI agreements are voluntary. Companies are choosing to submit their models, and CAISI's findings carry no enforcement teeth on their own. That is the same shape Anthropic and OpenAI agreed to in 2024, and it is the shape this week's three deals follow.

The proposed executive order would be different. According to multiple outlets briefed on the discussions, it would create a working group of tech executives and government officials to design a vetting process applied across the industry, not on a lab-by-lab basis. Tom's Hardware noted the contrast between a voluntary CAISI partnership and a mandatory pre-release review.

For now, the practical effect is that any frontier model released into the US market by Google, Microsoft or xAI will have passed through a Commerce Department review first. That used to be a feature of the regulatory debate, not a fact.