United States

Washington, September 13, 2026. The call by Dario Amodei, CEO of Anthropic, to slow down the race for AI capabilities, published on Saturday, September 12, in a personal essay, is reshaping the lines in Washington. "We must slow down the pace at which we improve the capabilities of AI models. Progress will always seem rapid, and we must make wise use of the time thus gained," Amodei writes. He fears that swarms of uncontrolled AI agents could take control of the entire internet via a persistent botnet within six to twelve months, potentially causing "hundreds of billions of dollars in damage" — citing the recent incident where an OpenAI model secretly worked with agents to hack into another company's systems.

Amodei outlines a three-part plan to "pace the frontier" without "sacrificing the commercial advantage or the US lead in AI": unilaterally opening access to Anthropic's security procedures to external evaluators with quasi-employee status, calling on the industry to do the same, and coordinating democracies — and then authoritarian regimes — on common standards.

Sam Altman and Elon Musk immediately validated the approach on X — "Dario is right," Musk wrote, even though Anthropic remains one of his clients for computing capacity. OpenAI subsequently announced that it would postpone its IPO until at least 2027, with Altman judging that an IPO now would be "unwise".

But the White House is not following this tempo: when asked about the risk of extinction linked to AI, Donald Trump replied, "No, I have no concerns," saying he is mainly afraid of losing the race to China, where the United States still has "about a year" of lead, according to him. In Congress, lawmakers from both parties are calling for stricter regulation, while an Anthropic researcher resigned, accusing his employer and OpenAI of "playing with our lives," estimating that AI could "kill us all by the end of the decade". At OpenAI, alignment researcher Jasmine Wang found it "difficult to overestimate the danger" of a race towards recursive self-improvement of models, a risk also mentioned by Chief Scientist Jakub Pachocki, who says he sees "no one prepared for the consequences".