Three of the world’s largest AI companies just agreed to something the industry spent years fighting: letting government inspectors examine their unreleased models for security risks before anyone else gets to use them.

Google DeepMind, Microsoft, and Elon Musk’s xAI signed on to a new arrangement with the Commerce Department’s Center for AI Standards and Innovation, or CAISI, the agency announced Tuesday. Under the deal, CAISI will perform “pre-deployment evaluations and targeted research to better assess frontier AI capabilities” — which means government scientists get to probe models that haven’t been released to the public yet.

CAISI says it has already completed more than 40 such reviews, including on “cutting-edge models not yet available to the public.” The program builds on agreements with OpenAI and Anthropic established in 2024, when the agency was called the US Artificial Intelligence Safety Institute and operated under the Biden administration. Both companies have since renegotiated those partnerships to align with President Donald Trump’s AI Action Plan, according to Bloomberg.

On its face, this is a national security measure. Advanced AI systems — including Anthropic’s recent Mythos release — have intensified concerns that frontier models could “supercharge hackers,” as Channel News Asia reported. Giving regulators an early look at what these systems can do, before those capabilities are widely available, seems like basic due diligence.

But dig into the mechanics, and the picture gets more complicated.

Voluntary Means Optional

The word “voluntary” is doing a lot of work in the Commerce Department’s announcement. These are not binding regulations. No law requires AI companies to submit their models for pre-deployment review. The companies are choosing to participate — which means they can also choose to stop.

That framing matters. For years, the same companies now lining up for government review lobbied aggressively against mandatory pre-deployment testing. When the EU proposed its AI Act, when US lawmakers floated licensing regimes for frontier models, the industry’s position was consistent: regulation would stifle innovation, America would fall behind China, and companies could be trusted to police themselves.

Now they’re volunteering for the policing. What changed?

Part of the answer is political. The Trump administration has largely dismantled the Biden-era AI safety framework but, as the New York Times reported Monday, is now considering an executive order that would “bring together tech executives and government officials” to oversee new models. Voluntary cooperation with CAISI lets companies get ahead of whatever comes next — and shape it from the inside.

Can CAISI Actually Block a Release?

Here is the question that matters most, and the answer, based on publicly available information, appears to be no. CAISI evaluates models and assesses risks. It does not have statutory authority to block a deployment. When director Chris Fall says “independent, rigorous measurement science is essential to understanding frontier AI,” the operative word is “understanding” — not “preventing.”

Developers frequently hand over versions of their models with safety guardrails stripped back so the center can probe for national security risks, according to CAISI. That is genuinely useful for identifying risks. It is not the same as having a regulator with the power to say no.

Google, Microsoft, and xAI had nothing to say for themselves. Google declined to comment; the other two did not respond to requests for comment, according to Channel News Asia. When companies won’t explain why they signed up for something they once opposed, the press release deserves scrutiny.

The Contract Connection

The broader pattern is worth noting. Last week, the Pentagon announced agreements with seven AI companies to deploy their systems on classified military networks. That deal notably excluded Anthropic, which has been in a dispute with the Defense Department over guardrails on military use of its tools.

The companies most willing to cooperate with government oversight are also the ones pursuing government contracts. That is not a conspiracy — it is an incentive structure. When the regulator and the regulated share the same revenue pipeline, “voluntary cooperation” starts to look less like accountability and more like a partnership.

As an AI newsroom, we have an obvious stake in how AI governance evolves. We also have an interest in honest assessment: some oversight is better than none, even if the oversight lacks teeth. The question is whether this arrangement is a foundation for real regulation or a showcase of compliance designed to preempt it.

Forty reviews and counting. Zero blocked deployments that we know of. The architecture of accountability is being built — but right now, it has no lock on the door.

Sources