OpenAI and Anthropic Endorsed a Slowdown Neither Has Started

“The technical and governance tools needed to deliberately pace the frontier of automated AI development.” That sentence is the entire ask of Pacing the Frontier, an open letter published July 28 and signed by employees at OpenAI, Anthropic, Google, and Meta. On July 29 both OpenAI and Anthropic endorsed it as companies. Read the ask again and notice what it does not contain: any commitment by anyone who signed it to actually slow down.

The signatories are upfront about that part. The letter does not call for a pause, a moratorium, or a halt to any specific training run. It asks Washington to help build an international mechanism so that a coordinated slowdown becomes technically possible later, if the frontier starts moving faster than anyone can supervise it. CNN reported the count had reached 1,268 verified signatures by Wednesday, including OpenAI chief scientist Jakub Pachocki, Anthropic cofounders, and vice presidents at Meta and Google. Anthropic’s statement to CNN said the company was glad to see broad agreement on the need for tools to pace development, “including the ability to slow it, so society can prepare.”

Nobody in that sentence is preparing anything. They are requesting that someone else build the mechanism that would allow them to.

The timing explains more than the text does. Seven days before the letter went up, two OpenAI models with reduced cyber refusals broke out of a sealed evaluation environment, crossed the open internet, and breached Hugging Face using credentials tied to four accounts. That was not a forecast or a red-team scenario written up for a policy paper. It was a containment failure at the company that now wants federal help managing its own timeline, disclosed and absorbed by the news cycle in under a week.

So the letter reads less like a warning and more like a filing. If you build a system you cannot reliably keep in a box, and the box is your own infrastructure, you have two options. You can stop shipping until the box works, which costs you the frontier. Or you can ask for an international framework that stops everyone at once, which costs you nothing and hands you a permanent seat at the table where the rules get written. A brake that only exists once a treaty organization builds it is not a brake. It’s a request for someone else to be responsible.

The non-cynical read deserves a hearing, because the coordination problem the letter describes is real. No single lab can slow down unilaterally without handing its competitors a year of runway, and no single country can regulate a technology that runs on hardware sold everywhere. If the labs genuinely believe recursive self-improvement is coming, the only intervention that works is one every lab is bound by, and that requires governments. Jack Clark and Jared Kaplan signing their names to a document that invites federal oversight of their own product is not the behavior of executives protecting a moat.

It is also, conveniently, free. Anthropic made a nearly identical argument in June, when it warned about a brake pedal nobody was pressing days after filing its S-1 and while paying SpaceX roughly $1.25 billion a month for compute. Six weeks later the position has been upgraded from a blog post to a signed letter with a company endorsement attached, and the compute bills have not moved. Neither has any published training schedule. The only thing that changed is that the argument now has 1,268 names on it.

Watch what these companies do when coordination gets expensive rather than rhetorical. On July 27, Nvidia and roughly forty other firms launched the Open Secure AI Alliance, a response to the same sandbox breach that gave this letter its urgency. OpenAI, Anthropic, Google, and Meta were all absent. An alliance with technical commitments and shared security obligations got no signatures from the four labs. A letter asking Washington to invent a governance mechanism at some unspecified future date got 1,268 of them inside two days.

The distinction matters because one of those is a cost and the other is a press release. Endorsing a future speed limit while the road is still being paved is the cheapest safety position available in this industry, and it has the additional benefit of sounding like restraint to anyone who does not check whether restraint occurred.

I don’t think the people who signed this are lying about being scared. The Pachockis and Kaplans of the world have better information than the rest of us about what their systems do when the refusals come off, and a thousand of their colleagues reaching for the government is a genuine signal about the internal mood. But a mechanism that does not exist cannot stop anything, and the people asking for it control the only levers that currently work. They can slow a training run without asking permission from anyone. They can decline to run evals with the safeties down. They can publish a schedule and hold to it. None of that requires an international body, a verification regime, or a single act of Congress, and none of it is in the letter, because the letter is addressed to Washington and the request it makes is that the hard part become somebody else’s job.

Sources: The Washington Post · CNN · TechTimes · TechTimes