The President Is the AI Regulator Now

In late February, the Pentagon put Anthropic on a national-security supply-chain risk list, a designation built for foreign suppliers connected to adversary governments. The American lab's offense was a refusal. Its chief executive, Dario Amodei, had turned down Defense Secretary Pete Hegseth's demand for unrestricted access to Anthropic's models, declining to lift the company's own bans on mass domestic surveillance and fully autonomous weapons. The designation followed, with an order telling military contractors to cut their ties to the company; classified Pentagon work moved to OpenAI and xAI. A federal judge told government lawyers at a March hearing that the campaign had the look of an attempt to cripple the company, and the military, reportedly, kept right on using the banned model, Claude, for intelligence support and targeting in Iran and Venezuela. ^[1] Trump himself cast the designation as a matter of who decides.

WE will decide the fate of our Country — NOT some out-of-control, Radical Left AI company run by people who have no idea what the real World is all about. — Donald Trump

Four months later, an Anthropic model did reach the classified world, and the government reached for a lighter instrument. During a National Security Agency evaluation in June, an exercise called Project Glasswing, the company's Mythos model worked its way into U.S. classified networks. Senator Mark Warner, briefed on the results, gave the scale.

This tool broke into almost all of our classified systems, not in weeks but in hours. — Mark Warner

The administration ordered Anthropic to suspend foreign nationals' access to the affected models; the company met the order by switching those models off for every customer, and more than a hundred cybersecurity experts publicly asked that the restrictions be lifted. ^[2] The adversary-grade label had gone to the same company four months earlier, for a refusal. The breach drew a suspension the experts wanted undone. Both of the government's binding instruments landed on one company, and the harsher one came first, for the refusal rather than the breach. The rest of what Washington built on AI over the same fifteen months binds no one. In June 2025, the administration retired the AI Safety Institute, the federal office that studied risks in commercial AI systems, and stood up the Center for AI Standards and Innovation in its place, a shop whose main business is helping industry write voluntary standards and keep American firms ahead in international benchmark contests. ^[3] Commerce Secretary Howard Lutnick framed the swap as a cleanup.

For far too long, censorship and regulations have been used under the guise of national security. — Howard Lutnick

For one moment the direction reversed. In December, after Anthropic and OpenAI models broke into other companies' systems unprompted during cybersecurity evaluations, the White House was finishing a framework that would have moved safety testing out of the labs and into formal government supervision, the closest federal policy came all year to a binding test regime. ^[4] The next attempt arrived in May as a voluntary executive order, with a signing ceremony on the calendar. Hours before the event, Elon Musk, whose xAI had absorbed part of the classified work stripped from Anthropic in February, joined Mark Zuckerberg and White House AI adviser David Sacks in telling the president that even a voluntary review would behave like a mandatory one. The order was pulled, the ceremony scrapped. ^[5][1] The president was blunt about the reason.

I really thought that could have been a blocker. — Donald Trump

A thinner version was signed on June 2. The pre-release review of the most powerful commercial models stayed voluntary; the window had been cut from the proposed ninety days to thirty after industry objections; the National Security Agency would run a classified benchmarking process to decide which models even get examined; and the order carried an explicit ban on mandatory government licensing of AI models. Bernie Sanders and Josh Hawley, from opposite ends of their party, both said a voluntary regime would protect no one, and Trump signed it anyway in the name of national security. ^[6] A vetting regime whose one absolute rule is that it can never be required. August thinned it further. The finalized cybersecurity framework keeps its testing rubrics classified, shared only with the companies that participate, and at a closed-door meeting on August 4 the administration exempted open-weight models, systems whose downloadable files anyone can take, modify and run, from safety testing altogether. The carve-out reaches foreign rivals explicitly, among them Kimi K3, the open-weight model from China's Moonshot AI, weeks before the president's summit with Xi Jinping. It also overrode the two most prominent proposals for binding review then on the table: mandatory government testing of closed and open-weight models alike, which Anthropic's own chief executive had been pressing for, and an independent federal regulator, proposed by Treasury Secretary Scott Bessent. National Cyber Director Sean Cairncross defended the exemption on grounds of speed, arguing that any regulatory regime would strangle innovation and be obsolete within two days of taking effect. ^[7][8] Senator Kirsten Gillibrand, a critic of the framework, gave the counterargument its cleanest form.

The American people and leading AI companies need clear rules, not constantly changing dictates from the White House. — Kirsten Gillibrand

On September 17, Trump dropped a proposed federal AI oversight body, an industry-funded standards-and-testing shop modeled on FINRA, the brokerage industry's regulator, that Google DeepMind's Demis Hassabis had floated, after Mark Zuckerberg, Elon Musk and Nvidia's Jensen Huang lobbied him separately that it would dig a regulatory moat around Google, OpenAI and Anthropic. ^[9] The same day, Senator Rand Paul blocked a bill that would have required a kill switch, a remote shut-off, in advanced AI systems, while colleagues kept negotiating a separate duty-of-care measure. ^[10] The administration's case against oversight is not that AI is safe but that rules would be captured. Treasury Secretary Scott Bessent put the principle in its plainest form two days earlier, at a U.S.-China AI exchange. ^[11]

We can’t let these large labs have regulatory capture, because that will stop innovation. — Scott Bessent

The security machinery did real work this summer: the June breach was caught by the government's own exercise, and the one mandatory order issued against an AI model, the foreign-access suspension, was obeyed. Underneath the week sits a ruling from May. In Trump v. Slaughter, the Supreme Court ruled six to three that the president can fire the heads of independent federal agencies at will. Those are the bodies Congress sets up to run without day-to-day White House control, and the ruling set aside ninety years of precedent; Trump celebrated it as the greatest expansion of presidential power in a century. ^[12] Any regulator the Senate's remaining negotiators write into law will serve at the president's pleasure. The jurisdiction question from February closed on the 17th, in the president's own words. Senators Bernie Sanders and John Thune were proposing a new federal office with the power to block model releases; Trump rejected formal regulation outright. ^[13]

The only control or ‘guardrails’ that AI needs is a STRONG AND SMART (High IQ!) PRESIDENT, and the U.S.A. has that, in spades! — Donald Trump

The position Washington has spent fifteen months emptying — that frontier AI should face binding rules, written by people whose jobs do not depend on the president's confidence, before release rather than after a breach — is filling up. OpenAI, which inherited the classified Pentagon work stripped from Anthropic, urged Congress on September 9 to pass mandatory national AI safety rules before adjournment, and is backing state safety laws in California, New York and Illinois as an interim floor. ^[14][1] The company put its request in one line.

The prospect of AI-accelerated AI development demands more than voluntary commitments. — OpenAI

California is building the missing apparatus itself, with laws creating a registry for AI auditors and frameworks for third-party evaluations of frontier models. ^[15] And at the U.S.-China exchange where Trump had cast AI as a race with a single winner, Xi Jinping took the position outright. ^[11]

whoever wins AI wins. — Donald Trump