Something genuinely historic happened this month and almost nobody is celebrating it.

Every major AI lab is now under federal review before launch

On May 5, NIST's Center for AI Standards and Innovation announced agreements with Google DeepMind, Microsoft, and xAI to evaluate their frontier AI models before public release. That brings the total to five major labs under voluntary federal pre-deployment review. OpenAI and Anthropic signed similar deals back in 2024. Now the whole roster is in.

I think this is genuinely good policy. Not perfect. Not sufficient on its own. But good. The people who spent years screaming that AI was being released into the wild with zero accountability now have something real to point to. The question is whether they will actually acknowledge it.

What triggered this shift? A model called Anthropic's Mythos. Anthropic said it was too dangerous to release publicly because of its ability to find serious software vulnerabilities with autonomous precision. That spooked national security officials. And when national security officials get spooked, administrations move.

What CAISI actually does when it gets these models

Here is the part most coverage glosses over. CAISI evaluates models with their safety filters stripped away. The labs hand over versions of their systems that have reduced or removed safeguards so government evaluators can probe what the model is actually capable of at full power. NIH handles biosecurity risks. National labs handle chemical and nuclear proliferation. DoD and DHS handle cybersecurity.

That is not a rubber stamp. That is a real evaluation architecture.

“Independent, rigorous measurement science is essential to understanding frontier AI and its national security implications.”

— Chris Fall, CAISI Director

I remember watching the early debates about AI safety in 2023 and thinking: nobody in Washington actually understands what they are regulating. That has changed. The TRAINS Taskforce now pulls evaluators from across the federal government, each assigned to a specific risk domain. This is what serious institutional capacity looks like when it starts to form.

The strongest objection, and why it does not land

Critics will say these agreements are voluntary and CAISI cannot actually block a launch. That is true. CAISI has no veto power. But that framing misses what is actually being built here. The federal government now holds a written, multi-lab record of what frontier AI models can do when their safety filters are stripped away. That classified record inside NIST is the closest thing the country has to an AI capabilities baseline.

Voluntary frameworks are how most serious regulatory regimes start. The FDA did not spring fully formed from the earth. Oversight infrastructure gets built incrementally, and right now the US is building it faster than most people realize.

The bad edge: what is still broken and needs fixing now

Here is my honest criticism: CAISI still has not published what it is actually testing for. Devin Lynch, a former director for cyber policy at the White House Office of the National Cyber Director, put it plainly: capability assessments are only as good as the threat models behind them. CAISI needs to define and publish its standards, not just announce who it is testing with.

That opacity is unserious. Accountability without transparency is just theater. If the public cannot see the evaluation criteria, the whole exercise risks becoming a political credentialing process rather than a genuine safety check. The administration needs to fix this fast.

Would you trust a food safety agency that refused to publish its testing standards? Tell me that is a reasonable ask.

Even xAI signed. That tells you everything.

The most telling detail in this whole story is that xAI signed. Elon Musk's company has been a loud public skeptic of AI safety regulation. It signed anyway. Microsoft, the most exposed to government procurement, signed the longest-term agreement. Google DeepMind, with the largest international footprint, signed terms similar to its US-only competitors.

When the most regulation-hostile lab in the industry decides the political cost of staying out is too high, you know the center of gravity has shifted. That is not capitulation. That is a new normal forming in real time.

In August 2024, two labs were under federal pre-deployment review. In May 2026, every major American frontier AI developer is. That is not nothing. That is a structural change in how the most powerful technology in human history gets released. The critics who wanted oversight got it. Now the work is making it count.