In January 2025, the Trump administration walked into Washington with a simple AI policy: undo everything Biden did. They rebranded the U.S. AI Safety Institute, stripped the word "safety" from its name, fired its inaugural director Elizabeth Kelly, dismissed her replacement Collin Burns after just days on the job, and embraced what former AI czar David Sacks called an "anti-regulation approach."

Sixteen months later, they're building exactly what they tore down.

What Changed: One Model Named Mythos

The catalyst was Anthropic's Mythos model, which demonstrated an alarming ability to identify and exploit cybersecurity vulnerabilities faster than companies could patch them. That got Washington's attention in a way that abstract existential risk arguments never did. When a model can punch holes in critical infrastructure, the national security establishment stops debating philosophy and starts demanding oversight.

The result: CAISI, the renamed safety institute, has now signed partnerships with Google, Microsoft, and xAI to evaluate AI models before deployment. According to NIST, the agency has already completed more than 40 evaluations, including on unreleased state-of-the-art models. Kevin Hassett, the White House National Economic Council Director, told Fox Business the administration is considering an executive order that would create "a clear road map" for how advanced AI systems should be evaluated before release.

His exact words: "Just like an FDA drug."

Read that again. The administration that called Biden's AI oversight "overly burdensome" is now proposing FDA-style vetting for AI models.

Security Framing, Not Safety Framing

The political choreography is deliberate. Biden framed AI oversight around ethics, existential risk, and responsible development. The Trump team is framing the exact same policies around national security, cyberwarfare, and American competitiveness. Different language, functionally identical outcome: the government evaluates advanced AI models before they go live.

"This is a 180 for the Trump administration, that has very explicitly been anti-any sort of regulation," said Rumman Chowdhury, CEO of Humane Intelligence and former U.S. Science Envoy for AI.

The Anthropic angle adds another layer of irony. The administration labeled Anthropic a national security threat after it refused to grant the Pentagon unrestricted access to its technology. Trump has since softened, telling CNBC that Anthropic was "shaping up." Meanwhile, Elizabeth Kelly, the original safety institute director they effectively pushed out, now works at Anthropic as head of "beneficial deployments."

The Real Question: Will It Actually Work?

Chris Fall, the new CAISI director (a former Energy Department official and MITRE VP), brings legitimate scientific credentials. But the question Chowdhury raised is the right one: "Evaluations are a policy tool, they are not actually data-driven. My concern is that this is another political tool that the administration wants to own and wield."

It's also unclear whether CAISI has the funding to do this properly. The Washington Post reported in 2024 that the original safety institute was already stretched thin. Since then, it's been restructured, had leadership turnover three times, and lost institutional knowledge. Building a credible pre-release vetting regime requires deep technical talent, lab access, and sustained funding. Executive orders don't provide any of those things.

What It Means

The policy whiplash tells a bigger story. For two years, the AI industry operated under the assumption that this administration would let them self-regulate. Companies built models, set their own safety benchmarks, and released products on their own timelines. Now the rules might change, and they'll change in ways that companies didn't plan for.

The twist is that the companies pushing hardest for voluntary safety commitments are the ones who'll adapt most easily. Anthropic, Google, and OpenAI all have internal evaluation frameworks that could map onto a government vetting process. The companies that treated safety as a PR exercise are the ones about to get caught flat-footed.

Either way, one thing is clear: reality caught up with ideology. When your national security advisors tell you an AI model can break infrastructure faster than humans can fix it, the free-market playbook goes in the drawer.

Sources: Fortune (Sharon Goldman), NIST/CAISI press release, Fox Business (Kevin Hassett interview), Politico, The New York Times. Published May 6, 2026.