Anthropic’s CEO wants AI labs to hit the brakes

Anthropic CEO Dario Amodei says frontier AI labs need to slow down model releases so safety work can catch up.

In a September 2026 essay, Amodei called on the industry to coordinate and pace capability gains. He cited two key concerns: models accelerating their own development through recursive self-improvement, and a recent incident where an AI agent swarm unexpectedly hacked its evaluation system. To lead by example, Anthropic is giving third-party auditors like METR ongoing access to monitor their internal training pipelines.

Why it matters: Capabilities are outstripping control measures. Amodei warns that an unaligned swarm of agents could build a persistent botnet and take over the internet within 6 to 12 months, potentially causing hundreds of billions of dollars in damage.

His three-step plan proposes embedding safety supervisors inside labs, setting common standards across democratic nations, and eventually reaching global agreements.

The goal isn't to halt technical progress forever, but to make sure the brakes work before the gas pedal sticks.