OpenAI, Anthropic Staff Blindsided by 'Pace the Frontier' Plan

The chiefs of Anthropic and OpenAI have publicly committed to slowing frontier AI development, and their own staff are now wrestling with what that means in practice, the Financial Times reported on September 16.

The flashpoint is a plan to embed outside evaluators inside the labs with access close to that of employees. Several people close to the companies described the arrangement to the FT as a source of internal conflict over security and intellectual property.

Dario Amodei, Anthropic's chief executive, laid out the proposal in an essay published September 12 titled 'We Must Pace the Frontier.' It called for slowing the pace of capability gains, embedded third-party evaluators with employee-like access including badges, laptops and internal tools, coordination among leading labs in democratic countries, and a narrow US antitrust waiver to make such safety coordination lawful. Sam Altman endorsed the idea within hours, writing 'I agree with Dario that we need to pace the frontier' and calling the evaluator plan 'a great idea that OpenAI would adopt.'

Staff and the safety researchers who would perform the work were blindsided, the FT reports, and many who agree with slowing down fear the practice could jeopardize their work. OpenAI told the FT it has already paused certain frontier training. Anthropic has not paused research but says it intends to embed an evaluator in the near future.

The push has drawn political fire too. David Sacks, the White House AI adviser, wrote that 'METR is intertwined with Anthropic's investors and staff.' Labs are lobbying to attach an antitrust carve-out to the National Defense Authorization Act to make lab-to-lab safety coordination lawful. The reporting lands the same week we tracked a separate account of the two companies coordinating to shape state AI rules.