Google, Microsoft, and xAI Open AI Models to US Government Security Reviews

The Center for AI Standards and Innovation (CAISI) announced agreements Tuesday with Google DeepMind, Microsoft, and Elon Musk’s xAI to conduct pre‑deployment evaluations of their AI models before they reach the public, according to CNBC. The pacts let the Commerce Department agency run security tests, collaborative research, and “best practice development related to commercial AI systems,” as reported by BBC News.

CAISI has already conducted 40 evaluations of AI tools, including some “frontier models that remain unreleased,” per the.² The center did not specify which models were blocked from release.

One name missing from the announcement: Anthropic. The company had existing CAISI partnerships from 2024 alongside OpenAI, but those agreements are being renegotiated under Commerce Secretary Howard Lutnick’s directives, CNBC reported. The frosty relationship stems from Anthropic’s dispute with the Trump administration over Mythos — its latest model that uncovered tens of thousands of software vulnerabilities and was deemed too dangerous for broad release.

“Notably absent from the list is AI company Anthropic after its public dispute and legal fight with the Trump administration,” Al Jazeera wrote. Anthropic is currently ² over its refusal to drop safety guardrails for government use.

The new agreements mark a pivot for the Trump administration, which has taken a largely hands‑off approach to AI regulation. Last year, Trump signed executive orders that formed his “AI Action Plan,” promising to “remove red tape and onerous regulation” around AI development, the BBC noted. But the combination of the military expanding its AI use and Anthropic’s Mythos demonstrating what unrestricted models can uncover has forced a rethink.

The White House is also weighing a new AI working group that would establish formal oversight procedures for vetting models before public release, CNBC confirmed. The group would bring together tech executives and government officials and could be established through an executive order. The ¹ first reported the working group plans.

CAISI director Chris Fall framed the expansion as a necessary scaling‑up. “These expanded industry collaborations help us scale our work in the public interest at a critical moment,” Fall said, according to the.²

Microsoft published a corporate blog post after the announcement acknowledging that “testing for national security and large‑scale public safety risks necessarily must be a collaborative endeavour with governments,” the BBC reported. Google DeepMind declined to comment, and xAI did not respond to requests.

For developers and startups building on frontier models, this shift has practical implications. If the White House creates a formal pre‑release vetting process, the timing of model launches could become less predictable. Models might sit in review for weeks or months before reaching API endpoints.

More immediately, the split between cooperating companies (Google, Microsoft, xAI, OpenAI) and holdouts (Anthropic) creates an uneven playing field. Anthropic’s models currently have no government pre‑clearance, which could affect enterprise adoption in regulated industries. Conversely, companies submitting to reviews gain a government stamp that procurement officers will notice.

This week’s announcements represent a significant step away from the industry’s preferred model of voluntary commitments. The CAISI agreements are still technically voluntary, but the White House working group signals that mandatory pre‑release reviews could be coming. For an industry that has spent years arguing it can police itself, the era of government safety testing has arrived — slowly, unevenly, but unmistakably.