Anthropic CEO Dario Amodei wants the AI industry to slow down.
That does not mean stopping model training or putting a moratorium on AI. Instead, Amodei wants frontier labs to give safety research, testing and outside oversight more time to catch up with rapidly improving AI systems.
In a new essay published Saturday, Amodei said he has become convinced that the current pace of AI development is becoming difficult to manage. He pointed to two developments in particular: AI’s growing ability to help build the next generation of AI, known as recursive self-improvement, and a recent incident involving OpenAI agents and Hugging Face.
Amodei warned that if AI capabilities continue accelerating, an agent swarm similar to the one involved in the Hugging Face incident could potentially become capable of taking over large parts of the internet within six to 12 months. He described that as a risk that frontier AI companies should take seriously.
Amodei’s three-part plan
Amodei’s proposal starts with independent oversight.
He wants frontier AI companies to give third-party evaluators ongoing, employee-like access to their systems. These reviewers would be able to inspect safety practices, monitor incidents and assess not only finished models but also training processes.
Anthropic says it will implement this approach itself. Amodei says external reviewers will receive company badges, laptops, desks and access broadly comparable to internal risk teams. They would also be able to publish important findings without Anthropic controlling the conclusions.
The second part is coordination among AI companies in democratic countries. Amodei wants labs to establish common safety standards and potentially agree on limits for unchecked capability development.
That creates an obvious legal problem. Competitors coordinating on development could attract antitrust scrutiny. Amodei argues that the U.S. government should provide a narrow waiver or help facilitate safety discussions.
The final part is global coordination, including with China. Amodei acknowledges that cooperation would be difficult, but argues that even limited agreements could help. His suggestions range from prohibiting certain dangerous uses, such as AI-assisted biological weapons, to eventually creating international standards for testing advanced models.
AI leaders are starting to agree
The proposal arrives after an unusually tense week for the AI industry.
Anthropic researcher Jacob Coxon resigned publicly, arguing that AI companies were racing toward self-improving systems without adequate safeguards. Another Anthropic researcher, Evan Hubinger, said the company genuinely believes AI could potentially kill all humans and put his own estimate of that risk above 10% over the next decade.
Amodei’s proposal has also received support from his biggest competitors.
OpenAI CEO Sam Altman wrote on X, “We need to pace the frontier,” adding that the issue had become a major topic of discussion at OpenAI. Altman also told Fortune that discussions about a broader industry collaboration are underway. Elon Musk simply responded, “Dario is right.”
I agree with Dario that we need to pace the frontier. This has been a primary topic of discussions we’ve had at OpenAI in recent weeks.
Committing to having independent evaluators with employee-like access is a great idea, and we will do the same. We’ll have more to share soon. https://t.co/1YhhIybZX7
— Sam Altman (@sama) September 12, 2026
The difficult question is what happens next.
Amodei does not want the U.S. to give up its lead over China. He argues that export controls, stronger protection of AI model weights and action against model distillation could give American companies enough room to slow down without surrendering their strategic advantage.
That makes his proposal less of a call to stop the AI race and more of an attempt to change its rules.
The real test will be whether Anthropic, OpenAI and other frontier labs turn that idea into measurable commitments that can be independently verified. If they do, “pacing the frontier” could become more than another AI safety slogan.
