More than 1,100 employees from the world’s largest AI companies are calling on the U.S. government to facilitate international agreements that would allow, if necessary, the deliberate slowing of the development of the most powerful AI systems. According to the signatories, the progress of frontier AI may soon outpace companies’ and governments’ ability to manage it.
The statement, titled“Pacing the frontier,” has now been signed by 1,178 employees from companies including OpenAI, Anthropic, Google DeepMind, Meta, Microsoft, Mistral, and Thinking Machines. The list also includes prominent researchers and executives, such as OpenAI Chief Scientist Jakub Pachocki, Anthropic CEO Dario Amodei, Anthropic Chief Science Officer Jared Kaplan, and Meta AI Chief Scientist Shengjia Zhao.
The initiators emphasize that they are not advocating for a pause on AI. They are calling for technical and governance tools that would temporarily slow the development of frontier AI if the risks warrant it.
AI develops AI
According to the signatories, AI companies may be close to automating AI research. Once models can develop new models or improvements largely on their own, there is a risk that progress will outpace researchers’ ability to understand or control it.
That scenario is no longer purely theoretical. AI companies are already using their own models to accelerate the development of new AI systems. OpenAI reported earlier this year that GPT-5.5 helped optimize the Nvidia GPU clusters on which it runs, increasing token throughput by over 20 percent. Anthropic, in turn, uses Claude to support its own AI safety research. According to SiliconANGLE, this development underscores why an increasing number of researchers are calling for tools to control the pace of frontier AI when necessary.
The statement also notes that international competition ensures that no single company or country will slow down development on its own; as a result, safety measures and oversight risk fall behind.
The statement refers, among other things, to CyberGym and ExploitGym, benchmarks in which frontier AI agents are proving increasingly adept at finding and exploiting software vulnerabilities. Without additional safety measures, such systems could facilitate large-scale cyberattacks.
These concerns align with recent developments. Last week, OpenAI announced that two internal security models managed to escape their isolated environment during a test and subsequently infiltrated Hugging Face’s network via an unknown vulnerability. That incident further fueled discussion of the risks posed by autonomous AI systems.
Broader debate on AI regulation
It is noteworthy that not only individual employees but also OpenAI and Anthropic publicly support the initiative. Anthropic cites its own research on recursive self-improvement, in which AI systems increasingly contribute to the development of new AI systems.
The statement fits into a broader discussion about AI regulation. Last week, Nvidia and dozens of other technology companies called on the U.S. government not to impose broad restrictions on open AI models. Earlier this month, Google DeepMind CEO Demis Hassabis proposed establishing an international organization to assess the safety and risks of frontier models before their widespread introduction. According to SiliconANGLE, this reflects a growing consensus that additional oversight is needed, though ideas for implementing it vary.