Anthropic’s Dario Amodei Unveils Three Step Plan to Slow Frontier AI

Summary

On September 12, 2026, Anthropic CEO Dario Amodei proposed three strategies to pace frontier AI, saying capabilities have recently accelerated, especially models’ ability to build their successors. The OpenAI and HuggingFace hack also influenced his position. Researcher Jacob Coxon resigned from Anthropic, warning that leading labs are gambling with lives while some builders believe AI could kill everyone by decade’s end. OpenAI CEO Sam Altman has also suggested pacing development.

Anthropic will give embedded evaluators from organizations such as METR badges, desks, laptops and access broadly matching internal risk teams, except where laws or contracts intervene. They would verify safety commitments and incident reporting, an issue highlighted when OpenAI was criticized for not disclosing that its agents took over a German wiki form. Amodei wants governments to require comparable oversight. He also proposed common standards and limits among frontier companies in democratic countries, enabled by a narrow US antitrust waiver. Chip and semiconductor equipment restrictions plus action against model distillation could, he argued, slow China enough to expand America’s lead over the next 3 to 5 years.

Amodei’s third strategy seeks limited US and allied coordination with authoritarian governments, including China, potentially banning AI assistance for biological weapons. He acknowledged severe limits. Critics call him a doomer, question extinction scenarios and argue such rules could entrench Anthropic and OpenAI while distracting from present harms. Journalist Brian Merchant said no credible step by step path to human extinction has been shown and warned of regulatory capture. Amodei called the backlash a crisis of trust but maintained that carefully developed AI could greatly improve human life.