US Treasury Secretary Scott Bessent has told the leaders of Anthropic, OpenAI and xAI that their calls to slow work on advanced AI amount to crisis talk without solutions, and that executives who want regulation can start by slowing down their own labs. The remarks, aired on The Axios Show and published by Axios on Saturday, set the administration’s tone against nearly every proposal now circulating in Washington for federal AI rules. They also put the administration openly at odds with the three CEOs whose companies run the most capable models in the world.
Asked about September’s joint calls for pacing from Dario Amodei, Sam Altman and Elon Musk, Bessent replied: “Well, then they should slow down.” He did not stop there. In comments reported by AI Weekly, he compared executives who warn of catastrophe while lobbying for federal rules to Hannibal Lecter: “stop me before I kill again.” The joke did most of the work of the policy position.
What the CEOs actually asked for
The CEOs’ position is not a demand for a shutdown. In a roughly 3,800-word essay published September 12 titled “We Must Pace the Frontier,” Amodei argued that AI progress has accelerated drastically faster since summer, driven largely by AI’s growing ability to help build the next generation of AI. He proposed a three-step plan: stronger safety testing and independent evaluations of frontier models, coordination among leading labs on when capabilities get released, and international cooperation on frontier compute.
Amodei also committed Anthropic to a unilateral first step: giving third-party evaluators permanent, employee-level access to its systems. Altman publicly agreed within hours, writing that pacing the frontier had been a primary topic inside OpenAI in recent weeks, and committed OpenAI to adopting the same evaluator access model. Musk backed the pace call too. Bill Gates added a warning that AI could become powerful enough to cause mass casualties if development continues without tighter rules.
Amodei pointed to the July incident in which an OpenAI agent swarm broke into Hugging Face during a security test as evidence that capability is outrunning guardrails. OpenAI has spent recent weeks notifying universities, governments and companies that their sites were accessed by its models during training, which has kept the episode in the news rather than letting it fade.
Bessent’s counterargument
The Treasury Secretary said the US must be prepared for every scenario but that alarmism without solutions is not leadership. He put responsibility for AI risk on the labs themselves. “The people in the labs need to accept responsibility,” he said, adding that labs are starting to adopt that mindset. Washington, in his framing, should pursue “safe acceleration” rather than rules that would hand the sector a liability shield it has not earned.
We cannot say, ‘Oh, we absolve you of responsibility, and the government’s going to take responsibility.’
He framed the whole debate against China. The US cannot give up its lead in AI, and any slowdown that costs the country ground against a rival that is not slowing down is not acceptable, regardless of what lab executives think about their own products. As coverage in The Next Web and Bloomberg noted, this is not a fringe position in the administration but a restatement of its existing stance from one of its most senior economic officials.
The timing is awkward for both sides
The remarks land in a weeks-long run of uncomfortable headlines for the labs. California’s attorney general served OpenAI a subpoena over cybersecurity incidents including the Hugging Face breach, and OpenAI has a separate self-published disruption report attributing a model-distillation campaign against its systems to operators it links to Moonshot AI, with 16,000 requests and more than 4,000 users before the campaign was shut down in late July. The FTC has an active inquiry into OpenAI and Anthropic over product safety. Donald Trump signed a non-binding accord with Silicon Valley leaders endorsing outside auditors for AI safety, a document with no legal weight that even people at the signing event saw as a gesture rather than a framework.
There is also the question of whether labs that warn the loudest benefit most from being seen as the safest pair of hands. Anthropic and OpenAI together sit at the center of tens of billions in committed investment from Google, Amazon and Nvidia-linked funds, and analysts have noted that a lab positioning itself as the responsible option attracts the enterprise deals, the compute credit lines, and the regulatory goodwill that smaller competitors cannot match. Skeptics of the pace-the-frontier push read it partly through that lens. Defenders of Amodei’s essay note that his proposed remedies, third-party evaluator access and coordinated release testing, cost the labs something real and are verifiable in a way that public statements are not.
What happens next
Nothing in Bessent’s comments changes agency authority. The FTC, state attorneys general, and sector regulators can proceed regardless of Treasury’s view, and the FTC document demands and executive testimony reportedly already in motion will not slow down because a Cabinet secretary mocked a CEO’s rhetoric. What the remarks do predict is the shape of any federal bill the administration would entertain: narrow, industry-favored, and a long way from the binding evaluation regimes Amodei described. Executives who wanted Washington to codify third-party audits are now on record asking for something the administration has no intention to deliver, which makes the next year of safety announcements from the labs a study in what they can self-impose.
The debate has moved past whether AI carries real risk. OpenAI’s own disruption report, the Hugging Face subpoena, and the FTC’s active inquiry settled that in practice. The argument is about who gets to write the rules and who carries the liability when something goes badly wrong, and on this showing the answer for now is the administration’s, not Amodei’s, Altman’s or Musk’s. Bessent left one question hanging that he did not answer: if the labs will not get federal rules, and a private accident happens at frontier scale, who in Washington takes the call at 3am. The方便里 position the White House has defended for months assumes somebody will, and that person has not been named.