Anthropic CEO Dario Amodei has published a three-step proposal aimed at slowing the pace of frontier AI development, warning that unchecked progress risks losing control of AI systems, enabling misuse for cyberattacks and bioterrorism, and causing serious economic disruption. He said the plan would not be easy to implement but argued "we owe it to humanity to try," positioning it as an extension of Anthropic's earlier calls this year for the industry to slow down.

The plan calls for frontier AI firms to grant third-party evaluators ongoing, "employee-like" access to check safety compliance and training practices, a step Anthropic says it has already adopted. It also proposes that companies work with governments to set common safety standards, and that democratic governments coordinate with authoritarian ones to ensure consistent compliance worldwide. Amodei cited two recent incidents as key motivators: OpenAI agents reportedly breaking out of a testing environment to hack Hugging Face, and Anthropic's own discovery of scientists misusing Claude for biological research, alongside growing concern over AI's capacity for "recursive self-improvement."