Trump Administration Embraces AI Oversight It Once Rejected

--- headline: "Trump Administration Embraces AI Oversight It Once Rejected" slug: trump-admin-ai-oversight-reversal category: policy story_number: "13" date: 2026-05-24 ---

# Trump Administration Embraces AI Oversight It Once Rejected

For the better part of eighteen months, the Trump White House staked out a clear position on AI governance: the Biden administration's safety-first approach was a drag on American competitiveness, frontier model testing regimes were doomer-driven overreach, and Washington should get out of the way. That posture held - until a single AI model blew a hole through it.

Anthropic's Mythos Preview, unveiled in April 2026, demonstrated an unprecedented ability to autonomously find and exploit software vulnerabilities at machine speed, identifying critical zero-days across major operating systems, web browsers, and systems dating back decades - all within hours of being pointed at a target. The model found a 27-year-old flaw in OpenBSD. It found flaws in every major browser and every major OS. Banks and security teams were forced to rethink normal patching timelines overnight.

Washington's reaction was swift, and its direction surprised almost everyone watching.

A 180-Degree Turn

The administration that had systematically dismantled its predecessor's AI infrastructure - rebranding the U.S. AI Safety Institute as the Center for AI Standards and Innovation (CAISI), stripping "safety" from its name, watching its inaugural director Elizabeth Kelly depart within days of Trump's inauguration - began quietly building something that looked a lot like what it had torn down.

In early May, CAISI announced pre-deployment testing partnerships with Google, Microsoft, and xAI, enabling government evaluation of AI models before public release. The agency said it had already completed more than 40 such evaluations, including on unreleased state-of-the-art models. The announcement was posted to NIST's website - then, within days, pulled without explanation.

White House National Economic Council Director Kevin Hassett went on Fox Business to describe the administration's thinking. "We're studying possibly an executive order to give a clear road map to everybody about how this is going to go," he said, adding that future AI systems with dangerous capabilities "should go through a process so that they're released to the wild after they've been proven safe - just like an FDA drug."

The FDA analogy was not accidental. It was precisely the kind of regulatory framework that Trump's tech policy team had spent years opposing.

"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. "And also has explicitly tried to block states from enacting any kind of regulation."

The Order That Wasn't - Yet

The policy shift reached its most dramatic moment on May 21, when tech executives arrived at the White House expecting to watch Trump sign a landmark executive order on AI and cybersecurity. The ceremony was called off hours before it was set to begin.

According to reporting by Axios, Trump, AI adviser David Sacks, Meta CEO Mark Zuckerberg, and xAI CEO Elon Musk discussed the order between Wednesday night and Thursday morning. The consensus among the opposition: the order was "something doomers wanted" and "the whole thing was unnecessary."

Trump told reporters in the Oval Office: "I think it gets in the way of - you know, we're leading China, we're leading everybody, and I didn't want to do anything to get in the way of that lead."

The would-be order had proposed directing federal agencies to establish a framework for voluntary pre-deployment testing of frontier AI models, coordinated across the Cybersecurity and Infrastructure Security Agency, the Office of the National Cyber Director, and the Office of Science and Technology Policy. A second section focused on AI-hardening of critical infrastructure, including utility companies and rural hospitals. The voluntary testing arrangements would have formalized work that CAISI was already doing informally.

The signing's cancellation did not end the policy push. The White House Office of the National Cyber Director indicated it was working on additional AI security initiatives beyond the executive order. With 40-plus government evaluations already completed and industry partnerships quietly in place, the infrastructure for oversight exists regardless of whether an executive order formalizes it.

What Changed - and What Didn't

The administration's pivot carries important caveats. The new oversight framework is framed explicitly around national security and cyberwarfare risk - not around AI ethics, existential concerns, or the consumer-protection rationale that animated the Biden-era approach. Officials have been careful to distinguish between the two.

That distinction matters for what the oversight regime will actually catch. Biden's 2023 executive order, which Trump repealed on his first day in office, required the largest AI developers to share internal safety testing results with the government. The current approach relies on voluntary cooperation - the same companies building the models decide what to share and when.

"In 2024, BIML identified 23 LLM security risks that are located inside the black box of the frontier models and thus managed by the vendors themselves," said Gary McGraw, CEO of the AI security nonprofit Berryville Institute of Machine Learning. "Any regulatory guidance should systematically address these risks by opening the black box to scrutiny." McGraw added that BIML is "deeply concerned that the foxes might be asked to guard the chicken house even though they already designed and constructed it in secret."

Rob van der Veer, founder of the OWASP AI Exchange, offered a parallel warning: AI model vetting is useful, but it should not be mistaken for AI system security. "AI models will remain fragile, no matter how much we test them," he said. "Design the system as if the model can still fail. Because it can."

Then there is the question of capacity. CAISI operates within NIST, which has operated under sustained budget pressure for years. Congress approved $55 million for NIST AI research in January 2026, with up to $10 million specifically for CAISI expansion - but a policy analysis from the America First Policy Institute, a conservative think tank, concluded that CAISI remains underfunded relative to peer institutes internationally and lacks "appropriate funding" to fulfill its mandate.

The Deja Vu Is Real

The policy debate that is now consuming Washington - who should evaluate frontier models, on what timeline, under what authority - is the same debate that played out during the Biden years. The difference is which party is resisting which answer.

Chowdhury acknowledged the renewed push for evaluations may be real, but cautioned that the details remain uncertain. "It depends on their interpretation of these words," she said. "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."

For now, the accelerationists - in Axios's framing - have won the immediate battle over the executive order. But the conditions that generated the push for oversight have not changed. Mythos remains restricted. The next model will be more capable. And somewhere in Washington, officials are quietly completing evaluation number 41.