A Hands-Off President Reaches for the Regulatory Lever
President Trump, who returned to office on a platform of stripping away government interference in artificial intelligence, is now considering doing the opposite — establishing a formal federal process to review new AI models before they are released to the public. The discussions, confirmed by U.S. officials who declined to be named given the sensitivity of the deliberations, center on a potential executive order that would convene an AI working group drawing together technology executives and senior government officials.
Among the options under consideration is a pre-release review mechanism that would give the government first access to powerful new AI systems while stopping short of blocking their deployment. In meetings last week, White House officials briefed executives from Anthropic, Google, and OpenAI on the outlines of those plans. The speed and seriousness of the discussions represent a notable break from Mr. Trump's public posture on AI, which as recently as last summer he characterized as a "beautiful baby" that must be allowed to thrive free of political interference and "foolish rules."
"The technology is moving extremely fast, and there are few formal procedures, but they also don't want to overregulate. It's a tricky balance."
How Mythos Changed the Calculus
The administration's posture began shifting in earnest last month after Anthropic announced a new AI model called Mythos. The company described the system as so proficient at identifying security vulnerabilities in software that its release could trigger a cybersecurity reckoning — and declined to make it publicly available on those grounds. That announcement landed in Washington with unusual force.
White House officials began discussing how to prevent the political exposure that would follow if a devastating AI-enabled cyberattack occurred with no government review process in place. But the deliberations are not purely defensive. Officials are also assessing whether frontier models like Mythos could yield cyber capabilities of direct value to the Pentagon and U.S. intelligence agencies. The National Security Agency has already used Mythos internally to probe vulnerabilities in U.S. government software — a fact that underlines both the model's power and the government's existing entanglement with it.
The review framework under discussion draws partial inspiration from the United Kingdom, which has assigned multiple government bodies to assess AI models against defined safety standards before they reach consumers. The U.S. version would likely involve a working group drawing on expertise from the NSA, the White House Office of the National Cyber Director, and the Director of National Intelligence. The Biden-era Center for AI Standards and Innovation — an agency built specifically to vet AI models shared voluntarily with the government — could also be rehabilitated for the role, having been largely sidelined under the current administration.
Silicon Valley Is Not Speaking with One Voice
The prospect of pre-release government review has not produced consensus among the companies most directly affected. Some executives have told White House officials that any formal oversight regime risks slowing American AI innovation at a moment when the competitive race with China demands maximum speed. Others have been more accommodating, recognizing that a government with early access to frontier models may be a more cooperative one. The companies do not agree among themselves on what U.S. AI governance should look like, and the administration's deliberations have reflected that fragmentation.
Bipartisan public concern may be pushing in the other direction. A Pew Research Center survey from last year found that 50 percent of Republicans and 51 percent of Democrats described themselves as more concerned than excited about the expanding role of AI in daily life. That near-identical split suggests the political cost of inaction could exceed the cost of intervention — a calculation that has not been lost on an administration sensitive to the next news cycle.
A $200 Million Standoff Complicates Everything
The path toward any new AI governance framework runs directly through one of the most consequential feuds in the industry: the ongoing battle between Anthropic and the Department of Defense. The two sides became embroiled in a dispute this year over a $200 million contract centered on how the military should integrate AI into warfare operations. When negotiations collapsed, the Pentagon cut off U.S. government access to Anthropic's technology in March. Anthropic has since filed suit against the government.
The fallout has been felt across military and intelligence agencies that had built operational workflows around Anthropic's tools. In a notable carve-out, however, Anthropic's AI continues to power a system known as Maven — the military's platform for analyzing intelligence and recommending airstrike targets in the war in Iran. The coexistence of the lawsuit and the ongoing Maven deployment captures the bizarre and unresolved nature of the relationship between the government and the company it is simultaneously suing and relying upon.
White House Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent, who have stepped in to fill the AI policy vacuum left by the departure of AI czar David Sacks in March, hosted a meeting with Anthropic CEO Dario Amodei last month. Both sides described the conversation as productive. Whether that thaw translates into a workable policy structure remains to be seen.
"The A.I. future is not going to be won by hand-wringing about safety. It will be won by building."
An Executive Order Without a Clear Author
Any executive order establishing a pre-release AI review regime would need to navigate competing institutional claims. No single federal agency currently holds authority over all government cybersecurity, which means the working group's first task would be sorting out which bodies — the NSA, the Office of the National Cyber Director, the Director of National Intelligence — should carry the review mandate. That institutional question is not trivial, and history suggests such turf negotiations can quietly bury policy proposals with broad support.
A White House official pushed back on the New York Times reporting, characterizing discussions of a potential executive order as speculation and noting that Mr. Trump would announce any policy himself. That denial does not resolve the underlying tension: an administration that spent its first months dismantling AI oversight is now being driven by events — a powerful unreleased model, a bipartisan political consensus forming around concern, and a high-stakes institutional dispute — toward exactly the kind of formal government engagement it once dismissed as foolish. How far that drift goes, and how fast, may define the next chapter of U.S. AI governance.
Any formal vetting regime would represent a direct contradiction of Vice President JD Vance's remarks at the Paris AI Summit last year, where he warned international officials that excessive regulation of AI could destroy a transformative industry before it fully emerged. Vance's position was framed as a rebuke of the European approach to AI governance. Squaring a new U.S. pre-release review process with that public record — without producing an obvious reversal narrative — is one of the political challenges officials must manage as these deliberations continue.