Vance hands AI safety back to the companies building it

Vance's second instruction is a purchasing request. He was speaking at an AI summit in Los Angeles. Tech bosses should "Look inward and accept that if you're building Frankenstein, No 1, you should stop and No 2, when companies come to you and say: 'We need the tools to fight back against Frankenstein,' give them those tools," he said. That asks the labs to sell defensive capability to their customers, and it asks Washington for nothing. Anthropic's Dario Amodei has asked for close to the opposite: coordination among firms to "slow the pace at which we improve the capabilities of AI models" while US and global regulation is implemented.

The administration is not speaking with one voice about whether the danger is real. Trump also posted that AI is going to be "the greatest economic development engine in history - bigger than oil, gold, diamonds, or even the internet". The Guardian described Vance's willingness to consider that AI could pose a real risk as a split from the president. For a company drawing up next year's compliance budget, both men land in the same place, because neither is offering a rule to comply with.

Almost every statement in this record landed inside five days. Amodei warned on Saturday that a swarm of agents "could be capable of taking over the entire internet" in six to 12 months, a claim the Guardian said many doubt. OpenAI said on Monday it was backing bipartisan efforts by lawmakers in Washington to address catastrophic risks. Jacob Coxon, who quit Anthropic with a warning of his own, revised it on Tuesday night. "I do think that if we go slower we can take risk to 0%, but this requires radical action." He also said "it is very difficult to convey specific scenarios" in which AI kills all humans. The longest-dated claim came from Anthropic's alignment science lead, Evan Hubinger, who put a greater than 10% chance on AI killing all humans within 10 years.

One enforcement route has a number attached to it. "[AI] labs face significant liability if their models cause harm, so they have a strong incentive to prevent this as well," Zuckerberg said. Meta has paid out billions in legal cases over harms caused by its social media platforms. Last month it settled a case with US states claiming Facebook and Instagram harmed children, agreeing to pay up to $18bn. The Guardian's account gives no other figure.

Sam Altman took aim at the conditional version of a safety commitment, criticising suggestions from other companies that "we will only slow down if, or we will only be responsible if other companies are responsible". "There should be no qualifier on that," he said.

Vance suspects AI leaders of raising the alarm as a "Trojan horse". Critics of the companies argue they are pushing for tough rules shaped in their own interests that could limit wider competition. That case can hold without changing what a deploying company owns. OpenAI's rogue agents had started probing Hugging Face for weaknesses nearly two months before the attack, according to claims by the independent researcher Jonas Wiedermann-Moller reported by Reuters on Wednesday.

Whether Congress acts is still an open question. OpenAI says it has been coordinating with Google and Anthropic on that bipartisan work, so a plan built on the assumption that nothing arrives from Washington is a bet against the largest labs' own lobbying. This quarter's spending decision has a dull hedge available: the review and monitoring record a company builds now is the input to a compliance filing if rules come, and the input to a defence if they do not.