US secures access to Big Tech AI models for security testing

Microsoft, Google and xAI have agreed to give the US government early access to their latest artificial intelligence models, as Washington moves to get ahead of potential security risks tied to rapidly advancing systems.

The arrangement, announced Tuesday by the Department of Commerce’s Center for AI Standards and Innovation (CAISI), allows federal researchers to test new models before they are publicly deployed and assess how they might be misused.

The timing reflects growing concern in Washington about what the newest generation of AI can do in the wrong hands. Much of the urgency has been driven by systems like Mythos, recently unveiled by Anthropic, which officials worry could significantly expand the capabilities available to hackers.

Under the agreement, government scientists will be able to run controlled evaluations designed to surface unexpected behaviour and identify vulnerabilities. Microsoft said it would work with federal teams to test its models “in ways that probe unexpected behaviors”, while also developing shared datasets and workflows for ongoing assessments. The company noted it has already signed a similar deal with the UK’s AI Security Institute.

The move follows through on a pledge made by the administration of Donald Trump in July to deepen cooperation with tech firms on screening AI systems for “national security risks”.

Inside government, CAISI is emerging as the central hub for that effort. The agency said it has already completed more than 40 evaluations, including tests on cutting-edge models not yet released to the public. In some cases, developers provide versions of their systems with safety guardrails reduced, allowing researchers to more fully probe potential risks.

“Independent, rigorous measurement science is essential to understanding frontier AI and its national security implications,” CAISI director Chris Fall said.

The new agreements build on earlier cooperation under Joe Biden, when CAISI operated under a different name and focused on voluntary safety standards and testing frameworks in partnership with companies like OpenAI and Anthropic.

At the same time, the government’s approach is widening beyond testing into deployment. The Pentagon recently struck a separate deal with a group of major technology firms — including Google, Microsoft, Amazon Web Services, Nvidia, OpenAI, Reflection and SpaceX — to integrate AI tools into classified military networks, aimed at supporting “warfighter decision-making in complex operational environments”.

Notably absent from that list is Anthropic, which has been in a public dispute with the Trump administration over the ethics of using AI in military contexts.

Market reaction to the latest announcement was muted. Microsoft shares edged down 0.6 percent in midday trading, while Alphabet rose 1.3 percent. xAI, which is privately held, does not trade publicly.