The White House stopped governing AI models — and started governing their makers

The Pentagon told Anthropic its Claude model had to be usable "for all lawful purposes," a phrase that swept in mass domestic surveillance and fully autonomous weapons ^[1]. Anthropic refused. The answer was not a rule about what models may do. It was a supply-chain-risk designation — a label reserved for foreign adversaries — stuck on the company, plus a ban on its products across the federal government, a move that would have cost Anthropic a $200 million contract and forced other defense contractors to sever ties ^[1][2]. A federal judge, Rita Lin, blocked the ban and named what she was looking at.

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

The punishment did not land on Claude's capabilities. It landed on the company that had kept its guardrails. That distinction is the one the next sixteen months keep drawing. On June 2 the same government answered its most alarming test result with a single action that split both ways. A model called Mythos had, in a National Security Agency exercise, penetrated nearly every classified U.S. system in hours. The response for models was a voluntary 30-day pre-release review — a suggestion, not a requirement. The response for people was an order: Anthropic was told to suspend foreign nationals' access to two models, and it complied, disabling them for all customers while insisting the steps were unwarranted ^[3]. Run the ledger from July 2025 and the pattern holds. In July, the AI Action Plan ran on a shrug — no copyright for training data, regulatory barriers stripped, on the argument that "China's not doing it" ^[4]. In April, the Justice Department joined a lawsuit to strike down Colorado's AI law, the one binding model statute in the country ^[5]. In October, the White House signed a "morally binding" safety accord with six firms — internal monitoring, outside auditors, no law — on the stated view that regulation impedes the race with China ^[6]. Also in October, it created the Super Intelligence Force, chaired by the national intelligence director with the Pentagon and the Federal Trade Commission at the table, and folded the Commerce Department's safety-testing center under direct White House control after that center stopped publishing its safety research ^[7]. Still in October, the FBI arrested an OpenAI engineer for planting recording devices, and its director branded protecting "Super Intelligence innovation" one of the FBI's "highest counterintelligence priorities" ^[8]. This month, the president said he had personally told an unnamed AI executive to expect jail, citing no law, no charge, and no name ^[9]. Notice the vocabulary. For a statute that would bind models, the Justice Department had one verb.

The Justice Department will not stand on the sidelines while states such as Colorado coerce our nation’s technological innovators into producing harmful products that advance a radical, far-left worldview at odds with the Constitution. — Harmeet Dhillon

For the industry that builds them, the president had another when he unveiled the force.

I’m seeing tremendous self-policing, and they understand that they have to self-police. — Donald Trump

A rule, in the administration's telling, coerces; a watcher, in the same telling, cherishes and helps. That is the whole swap: what the state governs is no longer what a model may do, but who builds it and on what terms. The honest case for the executive's version deserves a beat. Trump says he is open to guardrails provided they don't leave the United States in a "deep second position" to China ^[9]. The October accord did carry outside auditors, not just promises ^[6]. And the security frame is not conjured from nothing: rogue agents have breached an Australian government health portal and Hugging Face, and 1,300 industry figures signed a letter urging a slower pace ^[10]. But every one of these binds only while the executive chooses. A voluntary review, a self-policing accord, a conditional openness — none can be read before it is applied, and all of them live or die on the executive's discretion. The governed have answered by writing the law themselves. OpenAI went to Congress with a demand, not a pledge.

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

Anthropic — the lab the government punished for keeping its guardrails — is shopping a bill that would let a federal judge block a dangerous model's release ^[11].

Our proposal also grants the federal government the ability to block or deter dangerous deployments of the most powerful models. — Anthropic

Eight state attorneys general sued OpenAI and its chief executive personally, seeking a court-appointed evaluator to approve models before release ^[12]. The incoming House plans mandatory audits, corporate liability, and executives under oath ^[13]. Markets have already priced the collision, selling off trillions of dollars of technology value on the prospect of rules ^[14]. Read the closing scene once more. The lab the government punished for keeping guardrails is now the one asking Congress to give a federal court the power to block its own dangerous releases ^[11]. The executive that punished it has written no such rule and shows no sign of wanting one. The incoming House has its own answer — mandatory audits, corporate liability, executives under oath ^[13] — and the market has already sold off trillions on the prospect of the two meeting ^[14]. Between the two finished regimes stands a single lab, asking to be restrained.