U.S. to Vet Google, Microsoft and xAI Models Before Release

Google, Microsoft and xAI have agreed to provide the Commerce Department pre-release access to their frontier AI models so federal officials can test them for national security risks. The Commerce Department’s Center for AI Standards and Innovation announced the agreements on Tuesday.

Chris Fall, director of the center, described the effort as “independent, rigorous measurement science is essential to understanding frontier AI and its national security implications.” The department said the access will allow independent measurement and testing of advanced systems before they are widely released.

Officials expect the testing to produce objective performance data and to help the government evaluate specific risks such as automated discovery of software vulnerabilities, disinformation capabilities and other forms of misuse. The companies have not disclosed which models will be included or the detailed terms of access to the center.

White House officials have discussed issuing an executive order to formalize pre-release reviews of advanced AI systems and creating a working group to examine models before public rollout. Those talks were prompted in part by concern about models that can identify software weaknesses and other potential misuse risks.

Anthropic’s Claude Mythos model drew attention after company testing showed it could find weaknesses in software and cybersecurity defenses. Anthropic limited Mythos access to a set of startups and organizations rather than a broad public release. A browser vendor reported using the model during testing to find and patch 271 vulnerabilities in its product.

The Defense Department designated Anthropic as a supply chain risk to national security earlier this year. A federal appeals court declined to pause that designation while a related lawsuit proceeds. Administration officials have weighed whether to resume a partnership with Anthropic under new terms.

An independent developer published an open-source reconstruction of Mythos on a public code site. The project’s readme states it is speculative and not affiliated with Anthropic. The reconstruction drew significant attention and thousands of stars on the hosting platform. A prediction market tracking AI deployment assigned a low probability to a broad public release of Mythos by the end of June.

The agreements with Google, Microsoft and xAI come amid changes in federal policy. Since returning to office in 2025, the administration revoked a 2023 executive order that had required developers to share certain safety test results with the government. The administration has proposed a national framework to set AI standards without creating a new regulator and has challenged some state-level AI rules.

Department officials described the pre-release access program as part of ongoing engagement with industry to better measure model capabilities ahead of broader release and to scale assessments at a time when officials are focused on potential threats from powerful models.