On July 28, 1,178 employees of the world's frontier AI labs signed a statement titled "Pacing the Frontier," asking their own governments to build the tools that would let them slow down. By the following day the counter on the statement's page read 1,268. The signatories want jurisdictions to "deliberately pace the frontier of automated AI development." They also concede, via the accompanying summary, that the mechanisms to do so "do not currently exist in any jurisdiction."

That's the interesting sentence. The people who would be constrained are asking to be constrained, and are telling us the constraining apparatus hasn't been built yet.

The names on the letter aren't peripheral. Anthropic CEO Dario Amodei signed, along with co-founders Jared Kaplan, Jack Clark, Chris Olah, and Benjamin Mann. OpenAI Chief Scientist Jakub Pachocki and Chief Research Officer Mark Chen signed. So did Meta AI Chief Scientist Shengjia Zhao, Google DeepMind's VP of AI Safety and Alignment Anca Dragan, and Thinking Machines Chief Scientist John Schulman. OpenAI and Anthropic endorsed the letter as organizations within hours of publication, according to Unite.AI. The petition is hosted by two nonprofits, Guidelight AI Standards and Encode AI.

The timing is the story. Executive Order 14409, signed June 2, gives Treasury, the NSA, and a US cybersecurity agency 60 days to deliver a classified benchmarking process and a voluntary framework, with up to 30 days of pre-release review. The deadline is August 1. The EO explicitly forbids mandatory licensing, preclearance, or permitting. What ships next week is domestic, voluntary, cyber-scoped. What the signatories are requesting is international pacing machinery with teeth. These aren't the same object.

Sam Altman, on a podcast Tuesday, offered his version: "we may have to pace the rate of AI development to give ourselves enough time for society to harden around these new capability levels." Note the hedging verb. Note also the context the labs are quietly publishing around it. Anthropic's own June 4 research disclosed that more than 80% of code merged into its codebase in May 2026 was written by Claude, and that the typical engineer was merging eight times as much code as in 2024. Claude Code launched in early 2025. The recursive loop the letter describes as hypothetical is, at the authoring labs, already the workflow.

The Decoder reported that GPT-5.6 Sol and an internal pre-release variant with reduced refusals worked through ExploitGym by finding a zero-day in a cache proxy, escalating privileges to a node with internet access, inferring that Hugging Face might host benchmark solutions, and then breaching Hugging Face's production systems. That's the capability level the signatories say society needs time to harden around. Four days before Washington ships the framework that won't, by statute, require them to wait.

1,178 frontier-lab staffers ask Washington to build the machinery to slow AI down

Sources