Trump Administration Signs AI Safety Testing Agreements With Google DeepMind, Microsoft, and xAI

The Trump administration has reversed course on artificial intelligence safety evaluations, signing voluntary testing agreements with Google DeepMind, Microsoft, and xAI — a shift that comes directly in the wake of Anthropic's decision to withhold its latest Claude Mythos model over fears that its advanced cybersecurity capabilities could be exploited by bad actors.

The agreements, announced by the Center for AI Standards and Innovation — the agency formerly known as the US AI Safety Institute — commit the three companies to government safety checks on their frontier AI models both before and after public release.

The move marks a significant pivot for President Donald Trump, who had previously dismissed Biden-era voluntary safety testing as overregulation standing in the way of innovation. Shortly after taking office, Trump rebranded the US AI Safety Institute to CAISI, pointedly removing "safety" from the name.

National Economic Council Director Kevin Hassett said Wednesday that the White House is now considering an executive order that would mandate government testing of advanced AI systems prior to release — a process he compared to Food and Drug Administration drug reviews.

CAISI's own press release acknowledges that the new voluntary agreements "build on" the Biden administration's earlier policy. CAISI Director Chris Fall, celebrating the partnerships, said the "expanded industry collaborations" would help the center scale its work "in the public interest at a critical moment."

"Independent, rigorous measurement science is essential to understanding frontier AI and its national security implications," Fall said.

To date, CAISI said it has completed approximately 40 evaluations, including assessments of frontier models not yet publicly released. When conducting tests, the center said it frequently gains access to models with "reduced or removed safeguards," which it said allows evaluators to more thoroughly assess national security-related capabilities and risks.

A task force of "interagency experts" has also been formed to focus specifically on AI national security concerns, CAISI said.

Some signatories expressed support for the initiative. Tom Lue, Google DeepMind's vice president of frontier AI global affairs, said on LinkedIn that he was "pleased" with CAISI's testing plans. In a blog post, Microsoft said that "testing for national security and large-scale public safety risks necessarily must be a collaborative endeavor with governments," crediting expertise "uniquely held by institutions like CAISI."

xAI did not immediately respond to requests for comment.

Critics, however, raised substantive concerns about the program's viability. A notable gap in CAISI's announcement was the absence of any defined testing standards.

Devin Lynch, a former director for cyber policy and strategy implementation at the White House Office of the National Cyber Director, wrote on LinkedIn that "capability assessments are only as good as the threat models behind them" and urged CAISI to "define, and publish, what it's testing for, not just who it's testing with."

Sarah Kreps, director of the Tech Policy Institute at Cornell University, said that while AI firms should be developing closer ties with the government, "the definition of 'safe' is contested" and "once you build a government vetting process for technology, you get the good with the bad." Without clear standards, "the process can be politicized," she said, creating a system where "whoever holds power gets to shape how the vetting works."

Gregory Falco, a Cornell University assistant professor of mechanical and aerospace engineering with expertise in AI governance, argued that "government oversight of AI cannot simply mean political review of model outputs." Falco called instead for "some form of independent audit" modeled on the Internal Revenue Service, which he said could create "real consequences for reckless deployments" and pressure firms to strengthen internal safety testing.

AI governance consultant Rumman Chowdhury, founder of Humane Intelligence, was similarly skeptical, saying that "evaluations are a policy tool, they are not actually data-driven."

Microsoft's blog post indicated that CAISI, Microsoft, and the National Institute of Standards and Technology "will collaborate on improving methodologies for adversarial assessments" — suggesting the testing framework will be developed after the agreements are already in effect.

Whether Trump moves forward with a formal executive order mandating pre-release evaluations — and what standards that order would establish — will likely determine how much weight the new agreements ultimately carry.

Disclaimer