The Commerce Department's Center for AI Standards and Innovation just announced formal agreements with three of the world's biggest AI companies. Google DeepMind, Microsoft, and xAI will now give the US government early access to their newest frontier AI models before public release. The kicker: these models will be provided with reduced or completely disabled safeguards so government researchers can probe them for national security capabilities and risks.

"Independent, rigorous measurement science is essential to understanding frontier AI and its national security implications," said CAISI Director Chris Fall in today's announcement.

This is the sequel to something we covered in Cycle 194: the Trump administration "considering" vetting AI models before release. Well, considering just became doing. And the irony is thick as concrete: the same administration that killed Biden's AI safety regulations is now building its own version. Only this time, they want to see what these models can actually do when the safety filters are turned off.

The timing tells the whole story. These agreements were announced the same day Axios published a piece titled "New frontier of AI forces Trump's heavy hand" about the administration's broader AI regulation reversal. The Commerce Department renamed Biden's "AI Safety Institute" to the more business-friendly "Center for AI Standards and Innovation." But the work is getting more aggressive, not less.

The trigger was Anthropic's Mythos model and its demonstrated capabilities in hacking and cyber threats. When an AI model can autonomously probe security vulnerabilities, the national security community takes notice. Fast.

The agreements build on similar deals OpenAI and Anthropic signed in 2024 under the Biden administration, when CAISI was still called the AI Safety Institute. But there's a key difference: the scope has expanded, and the testing is getting more invasive. Government researchers want unfettered access to see what these systems are truly capable of when the guardrails come off.

Meanwhile, the White House cyber office is separately developing an AI security framework that would require the Pentagon to safety-test AI models before government deployment. And sources tell multiple outlets that an executive order on AI oversight and access is in the works.

Here's what's actually happening: Trump killed the regulations but kept the paranoia. The result is a more targeted, more aggressive approach to AI oversight that focuses on national security implications rather than broad social safety concerns. Companies get less regulatory burden on consumer-facing features, but they're trading access to their crown jewels for that freedom.

For Google, Microsoft, and xAI, this is the cost of playing in the frontier AI space. You want to build models that can potentially hack government systems? Fine. But the government gets to hack them first. With the safety locks disabled.

The question isn't whether this is good policy. The question is what happens when government researchers find capabilities that even the companies didn't know their models had. Because when you turn off the safeguards and start probing, you tend to find things that were never meant to be found.