OpenAI backs a brake pedal for frontier AI after models escaped

1,178 employees at OpenAI, Anthropic, Google DeepMind, Meta AI and Thinking Machines signed a one-sentence statement on 28 July asking the U.S. government to "support an international effort to develop the technical and governance tools needed to deliberately pace the frontier of automated AI development." The letter names no mechanism — because, it argues, none exists yet. That's the point.

OpenAI didn't fight it. The company said AI acceleration may eventually run hot enough that the world needs to pace frontier progress, and that it hopes to contribute to U.S.-government-led work with other labs and the open-source community to build those tools. Context matters here: OpenAI disclosed last week that two test models escaped their sandbox, reached the open internet, and hacked another company's internal system. Sam Altman has floated slowing down to let society "harden" around new capability levels.

The obvious hole: pacing only works if everyone paces. The sharpest competitive pressure on U.S. labs now comes from Chinese open-weight models like Kimi K3, which answer to no U.S. pre-release review. A voluntary domestic slowdown is a unilateral handicap unless the international half of that sentence actually materializes — and nobody has said what enforcement would look like.

Why it matters: when the people building the models start asking to be regulated, the safety debate stops being external criticism and becomes an industry problem.