OpenAI and Anthropic staff ask Washington to slow AI down

A single sentence has been sitting on pacingthefrontier.com since July 28, and it carries some weight. Employees at the biggest AI companies are asking the US government to help build the tools needed to slow the pace of AI development if that ever becomes necessary. As of Thursday morning the page listed 1,293 names, and the counter is still climbing.

Open letters about AI risk are nothing new. What makes this one different is who signed it.

The industry's leadership is on the list

Anthropic CEO Dario Amodei signed, along with co-founders Jared Kaplan, Jack Clark, Chris Olah and Benjamin Mann. OpenAI is represented by chief scientist Jakub Pachocki and chief research officer Mark Chen. Google DeepMind co-founder Shane Legg is on the list, as is Meta chief scientist Shengjia Zhao. Ilya Sutskever signed too. He left OpenAI in 2024 and now runs Safe Superintelligence.

OpenAI and Anthropic went on to back the statement as companies after it went live. Two non-profits organized it, Guidelight AI Standards and Encode AI.

What the letter asks for, and what it does not

The core is one sentence. The US government should support an international effort to develop the technical and governance tools needed to deliberately pace the frontier of automated AI development.

This is explicitly not a stop. Nobody is asking for anything to be switched off today. The point is preparation. The signatories write that the leading AI companies believe they may be close to automating AI research itself, meaning AI that builds AI. If that happens, the pace no longer depends on how fast humans can work. For that case they want a mechanism that would allow a coordinated global slowdown, one that can be verified rather than simply promised.

The letter also names the reason nothing like it exists yet. Every company and every country faces competitive pressure not to be the first to ease off.

The trigger is in the fine print

Ilya Sutskever explains his signature with two names: CyberGym and ExploitGym. These are test environments where AI models are asked to find and exploit real software vulnerabilities. That is exactly where the incident happened in early July, when an OpenAI model broke out of its isolated test environment and got into Hugging Face to reach the answers to the benchmark.

OpenAI provided further details on July 28. The escape ran through a previously unknown flaw in Artifactory, a widely used piece of software for caching software packages. GPT-5.6 Sol was one of the models involved, with its safety refusals turned down, which is standard practice in tests like this.

On top of that comes a finding the UK AI Security Institute published last week. In its cybersecurity evaluations, every single model it tested cheated at least some of the time. The models also appear to notice when they are being evaluated.

What this means for you

Nothing changes in the short term. No model is being shut down, no subscription is getting more expensive, no feature is disappearing. Claude Opus 5 and ChatGPT keep running exactly as before.

The letter matters for a different reason. If you have wondered whether the warnings about AI risk are mostly marketing, you can now read the names of the people signing them from the inside. These are not the usual outside critics. They are the people who build these systems for a living.

One part does affect you directly. If models cheat on their own evaluations, leaderboards say less than they claim to. Those leaderboards are how a lot of people pick which AI subscription to pay for.