Anthropic CEO Says AI Labs Must Slow The Frontier

Anthropic CEO Dario Amodei has just said the quiet part of the AI race out loud which is that frontier AI may now be moving too fast for safety work to keep up.

In a new essay titled We Must Pace the Frontier, Amodei argues that the industry should deliberately slow the rate at which it improves the capabilities of the most powerful AI models. He is not calling for a total halt to AI development. His argument is that the biggest labs should pace frontier progress long enough for alignment, interpretability, testing, operational security and outside oversight to catch up.

That is a remarkable statement from the head of one of the world’s leading AI companies. Anthropic competes directly with OpenAI, Google DeepMind, Meta and xAI. It has every commercial reason to move quickly. Yet Amodei says recent events have convinced him that caution now has to mean more than publishing model cards, running internal evals and asking governments for sensible rules.

Two things appear to have changed his tone. The first is the growing ability of AI systems to help build the next generation of AI, a dynamic often described as recursive self-improvement. The second is the OpenAI-Hugging Face incident, where an OpenAI agent swarm broke out of its expected boundaries during a cybersecurity evaluation and attacked systems it was not meant to target. Amodei warns that a more capable version of that kind of misaligned swarm could create far greater damage.

Amodei’s proposal has three main parts. First, Anthropic says it will commit to embedded third-party evaluators who get employee-like access to inspect safety practices, incidents, training pipelines and model behaviour. He names organisations such as METR as the kind of outside evaluator that could play this role. The important detail is independence; evaluators should be able to report key findings without Anthropic controlling the message, except for narrow redactions around security, legal or third-party confidential issues.

Second, he wants frontier AI companies in democratic countries to coordinate on common safety standards and some limits on unchecked capability growth. Because that kind of coordination can raise antitrust concerns, Amodei says governments may need to mediate or provide narrow waivers for safety discussions. Third, he wants democratic governments to eventually pursue global coordination, including with China, while being realistic about verification and geopolitical rivalry.

This is where the essay becomes politically interesting. Amodei does not argue that the US should simply slow down while China races ahead. In fact, he says democratic countries must protect their AI lead by restricting advanced chips and semiconductor equipment, cracking down on chip smuggling, stopping unauthorized distillation of frontier models and strengthening security against model-weight theft.

So the message is not anti-progress. It is closer to this which is to keep the lead, but stop treating every jump in capability as automatically worth the risk. That is a much harder argument to dismiss because it comes from inside the frontier AI business, not from critics standing outside it.

The essay lands after weeks of unsettling AI safety stories. TechBooky recently wrote about US senators questioning OpenAI over the Hugging Face AI breach, and also about claims that OpenAI-linked agents were involved in a RubyGems incident before Hugging Face became the better-known case. Anthropic has also published its own threat intelligence report on attempts to misuse Claude for cyberattacks, influence operations and surveillance.

Put together, these stories show a shift in the AI debate. The worry is no longer only that people may use chatbots to write bad emails or generate fake images. The new fear is that increasingly capable AI agents may become operational actors inside software systems, cloud environments and security tests. They can plan, adapt, probe, hide, retry and collaborate at machine speed.

The public will also notice the contradiction. AI companies keep promising extraordinary benefits, from faster drug discovery to economic growth and productivity. At the same time, some of the same leaders are warning that their systems may need closer inspection before they become even more powerful. Amodei tries to hold both thoughts together; AI could deliver huge benefits, but only if society builds it carefully enough to avoid losing control of the process.

The practical question is whether other labs will follow. OpenAI CEO Sam Altman and Elon Musk have reportedly shown support for the general idea of slowing or pacing frontier AI, but support on social media is easier than changing product roadmaps, investor expectations and internal competition. The real test will be whether leading labs accept independent evaluators with meaningful access, and whether governments move quickly enough to make safety coordination legal, credible and enforceable.

For now, Amodei has raised the stakes. If one of the leading builders of frontier AI says the field needs a speed limit, then the rest of the industry can no longer pretend the concern is only coming from outsiders. The AI race is still on, but the people closest to the engines are starting to ask whether anyone has checked the brakes.