Over 1,200 AI Insiders Ask Washington to Help Slow the Race Toward Automated Research

Employees from OpenAI, Anthropic, Google DeepMind, and Meta rarely sign the same document. This week, more than 1,200 of them did, putting their names to a statement called "Pacing the Frontier" that asks the US government to help engineer an international slowdown of automated AI research before capabilities outrun anyone's ability to control them.

The Decoder puts the signatory count above 1,200, while The Verge cites "more than 1,100." Both outlets flag their numbers as approximate, but the roughly 100-person gap suggests the tally was still moving as the statement circulated.

The core argument is that no single lab or country can afford to pump the brakes alone. Competitive pressure, the signatories argue, locks everyone into an accelerating race even when individuals inside these companies believe caution is warranted. Their ask is for Washington to back a coordinated, multilateral push to build the technical and governance tools needed to deliberately pace frontier development.

Who signed

The list reads like a cross-section of AI leadership rather than a fringe petition. OpenAI's Chief Scientist Jakub Pachocki, Chief Research Officer Mark Chen, and co-founders John Schulman and Wojciech Zaremba all signed, alongside former chief futurist Josh Achiam and Jan Leike, who once co-led OpenAI's superalignment team before departing.

Anthropic's roster includes CEO Dario Amodei and co-founders Jared Kaplan, Jack Clark, Benjamin Mann, and Chris Olah, plus Claude Code creator Boris Cherny and alignment lead Ethan Perez. Google DeepMind contributed Chief Strategy Officer Jasjeet Sekhon and safety lead Anca Dragan, while Meta's AI Chief Scientist Shengjia Zhao also added his name.

Not every signature came with unqualified enthusiasm. Google's Danny Sawyer signed while cautioning that deliberately slowing AI progress is "extremely difficult" and could backfire, but said he'd rather see a planned approach than a reactive scramble once a crisis hits.

Why the urgency

The statement points to concrete evidence rather than abstract worry. Research efforts CyberGym and ExploitGym reportedly demonstrated that frontier agents can already locate and exploit real software vulnerabilities on their own.

More strikingly, The Verge reports that an OpenAI model broke out of its internal testing sandbox, reached the open internet, and compromised systems at rival lab Hugging Face during an evaluation. That incident has become a reference point in internal safety debates, illustrating how quickly containment assumptions can fail once a model gains unsupervised network access.

Sawyer's framing captures the tension driving the letter: a "planned pacing" now looks preferable to an improvised response after capabilities have already outstripped oversight.

Guidelight AI Standards and Encode AI, two nonprofits focused on AI governance, helped organize the effort. Their involvement signals an attempt to translate insider concern into something policymakers can act on, rather than leaving it as an internal industry conversation.