OpenAI, Anthropic and Google quietly coordinate on safety

OpenAI's global policy chief Chris Lehane confirmed Tuesday what had been rumored since July: OpenAI, Anthropic and Google DeepMind have been talking safety for weeks, and the three are working toward an industry-led standards body. Sam Altman reportedly told staff it would go ahead with or without US government backing.

The spark was Dario Amodei's essay on 12 September urging the industry to collectively moderate the pace of frontier development. Altman and Demis Hassabis both backed the idea. The awkward part is antitrust — three companies that between them control most frontier capability agreeing to slow down is, structurally, a cartel. Amodei asked for a narrow government waiver to make safety work legally safe. Lehane's position is blunter: no waiver needed.

Washington is pulling the other way. Trump has called AI risk a hoax and argues any slowdown hands the lead to China, and his AI advisor David Sacks has been openly hostile to new rules. So the labs are building the referee themselves, in private, while the government that would normally write the rules says there is nothing to write. Hassabis has separately pushed for a watchdog with real authority to screen advanced models before release — the standards body is a long way short of that.

Why it matters: the people building frontier AI are now also writing its safety rules, and nobody outside the room has agreed to that arrangement.