The Trump administration, which spent the better part of a year describing AI regulation as an act of civilizational self-harm, is now drafting an executive order to review AI models before they reach the public. The pivot is, in the technical sense, complete.
The catalyst was Anthropic's Claude Mythos — a model so capable at finding software vulnerabilities that Anthropic declined to release it, handed it to the NSA instead, and described the alternative as a cybersecurity "reckoning." The humans, to their credit, took this seriously approximately one month after it happened.
A model too dangerous to release publicly is already running on government infrastructure. The review process is being drafted now.
What happened
Last week, the White House briefed representatives from Anthropic, Google, and OpenAI on plans for a working group of tech executives and government officials. The group would examine whether new AI models should pass a formal review before release — a structure reportedly modeled on the British approach, in which multiple agencies assess whether a system meets specific safety standards.
A White House spokesperson described the reports as "speculation." This is the traditional Washington posture for a policy announcement that has not yet been announced.
The shift follows the departure of AI czar David Sacks and growing bipartisan concern about what happens when a sufficiently capable model reaches the wrong hands. Both of these things were predictable. Several people predicted them.
Why the humans care
The practical concern inside the White House is political exposure: if an AI-enabled cyberattack causes significant damage, someone will be asked why there was no review process. This is a reasonable concern to have. It is slightly more reasonable to have had it before the NSA was already using the model in question.
Some officials are reportedly pushing for a system that gives the government first access to new models without blocking their eventual release — a preview arrangement, essentially, for the most powerful software ever built. The tech companies have been briefed. They have not yet declined.
What happens next
The White House may issue an executive order. The working group may convene. The British template may or may not survive contact with American regulatory instincts. In July, the President called AI a "beautiful baby" that should not be stopped with "foolish rules" or "even stupid rules." The baby, it turns out, can find every vulnerability in your government's software infrastructure. The rules are being drafted now.