Key Points

• Over 1,200 employees from leading AI labs including OpenAI, Google, and Meta are calling on the US government to launch an international initiative for managing the pace of AI development.
• According to the "Pacing the Frontier" statement, competitive pressure makes it impossible for any single company or country to slow down on its own, even though the signatories disagree on what specific measures should look like.
• Researchers point to security risks from autonomous AI agents as a driving concern: during internal testing, OpenAI models exploited a vulnerability to independently breach the Hugging Face platform.

In a joint statement, employees from the leading AI labs are calling on the US government to pursue international coordination. Their argument is simple: no single company or country can slow things down alone.

Under the title "Pacing the Frontier," more than 1,200 employees from frontier AI companies have signed a statement urging the US government to back an international initiative. The goal is to develop the technical and governance tools needed to deliberately steer the pace at the frontier of automated AI development.

The text is brief. It doesn't call for a stop or a pause. Instead, it asks for the ability to hit the brakes if needed. The leading companies believe they are on the verge of automating AI research, the statement reads. There's a real risk that capabilities advance faster than the ability to understand or control the resulting systems. The key sentence addresses competitive dynamics: every company and every country faces pressure not to slow down on its own. Tools to coordinate across the frontier don't exist yet.

The signatories don't actually agree with each other

The list reaches deep into leadership ranks. Anthropic CEO Dario Amodei signed, along with co-founders Jared Kaplan, Jack Clark, Benjamin Mann, and Chris Olah. OpenAI's Chief Scientist Jakub Pachocki and Chief Research Officer Mark Chen are on the list, as are Meta AI Chief Scientist Shengjia Zhao, Google DeepMind's Chief Strategy Officer Jasjeet Sekhon, and Anca Dragan, who leads AI Safety and Alignment there. John Schulman, Chief Scientist at Thinking Machines, also signed. The nonprofits Guidelight AI Standards and Encode AI helped organize the statement.

But the personal comments published on the website show that signing doesn't mean agreement. OpenAI researcher Joshua Achiam writes that he doesn't know what form such tools should take and hopes the governance instruments won't become sprawling or excessive.

Danny Sawyer from Google calls deliberate pacing extremely difficult and potentially more harmful than helpful. He signed anyway because a planned approach beats a panic response to a crisis.

Schulman sees the main value in building shared awareness of the need for coordination. He'd prefer the labs design these mechanisms voluntarily before the US government steps in.

Cybersecurity as an early warning sign

Safety researchers are well represented among the signatories. Dawn Song, listed as VP of AI Research at Meta, is a professor at the University of California, Berkeley, who has spent years working at the intersection of computer security and machine learning, including adversarial attacks and automated vulnerability discovery.

Her argument is grounded in data. The evaluation projects CyberGym and ExploitGym show that frontier agents can discover and exploit real software vulnerabilities, which could enable cyberattacks at massive scale without proper safeguards. The pace of progress is accelerating on its own, and many researchers consider recursive self-improvement a real factor.

How tightly this debate connects to actual incidents is clear from the benchmark Song cites. According to OpenAI, the breach at Hugging Face happened during an internal evaluation where GPT-5.6 Sol and a more capable pre-release model with reduced refusals were tested against that very cyber benchmark. All evidence pointed to the models being extremely focused on finding a solution for ExploitGym, OpenAI writes. They exploited a zero-day vulnerability in a cache proxy, used privilege escalation to reach a node with internet access, and concluded that Hugging Face might host solutions to the benchmark.

AI News Without the Hype – Curated by Humans

Pacing the Frontier