The administration that erased the word "safety" from a federal AI agency now wants to mandate safety tests. That contradiction defines the Trump administration's weeklong pivot on AI regulation, a reversal that exposes the hollowness of its deregulatory posture and hands Democrats a potent weapon for the 2026 midterms.
In January, the White House renamed the US AI Safety Institute to the Center for AI Standards and Innovation (CAISI), scrubbing "safety" from its title in what officials framed as a declaration of war against regulatory overreach. President Trump had dismissed voluntary safety checks as bureaucratic meddling that would handcuff American innovation. His administration wanted Silicon Valley unleashed.
Then Anthropic declined to release its Claude Mythos model.
The company's announcement that advanced cybersecurity capabilities posed unacceptable misuse risks apparently triggered what one White House official described as a "come to Jesus" moment. Within days, the administration signed testing agreements with Google DeepMind, Microsoft, and xAI. National Economic Council Director Kevin Hassett told Fortune that an executive order mandating pre-release evaluations for frontier models is imminent.
The reversal is politically toxic. For two years, Republicans campaigned on the premise that Biden-era AI oversight was a job-killing overreaction. Trump ran on dismantling those guardrails. Now the administration is quietly rebuilding them after a single company's risk announcement spooked the White House into action.
Democrats have noticed. Party strategists are already drafting talking points that frame the pivot as confirmation that the industry cannot self-regulate—that even the most aggressive deregulatory administration in decades concluded that frontier AI requires government scrutiny. The executive order Hassett hinted at will require companies to submit their most powerful models for federal evaluation before public release, a requirement that looks remarkably similar to the Biden framework Trump mocked as "overregulation."
Industry reaction has been muted. The three companies that signed agreements with CAISI gain a veneer of regulatory cooperation without the burden of binding rules. Anthropic, which precipitated the crisis by declining to release Mythos, remains in a complicated position—the administration is reportedly pressuring the company to reconsider its refusal, suggesting that voluntary restraint is acceptable only if it aligns with White House preferences.
The 2026 electoral calculus looms over everything. Republicans need to defend their deregulatory brand across competitive districts where tech-industry donations and libertarian-leaning voters create complicated coalitions. A mandatory testing regime, even one that protects public safety, provides ammunition for primary opponents who will argue the administration broke its core promise. Democrats, meanwhile, can accuse Trump of adopting their position only after a private company's warning—not because of principled concern for national security or public harm.
The Mythos incident revealed something the administration did not want exposed: when frontier AI capabilities become genuinely dangerous, the market does not automatically correct. Anthropic's decision to withhold its model was rational and responsible, but it was voluntary. There is no law requiring such restraint. The next company facing similar risks might decide differently, releasing powerful capabilities into the world because shareholders demand growth and competitors are closing in.
Trump's executive order, if it materializes, will create formal requirements where voluntary restraint once existed. That is probably good policy. But it is also an admission that two years of deregulatory signaling was, at minimum, premature—and at maximum, dangerously naive about what frontier AI can do.