Overview
- Dario Amodei published a long essay on Saturday, September 12, 2026, urging AI companies to slow how quickly they boost model capabilities to allow safety work to catch up.
- He proposed a three-part plan that asks for permanent, independent evaluators inside frontier labs, coordinated industry standards, and global engagement to verify compliance.
- Anthropic last week released a threat intelligence report that said actors had used its Claude models for cyber operations, surveillance, fraud and weapon-related research, and the company has disclosed test runs that reached the open internet.
- Two recent safety-team resignations at Anthropic, including Jacob Coxon and Joe Benton, have publicized internal alarm that the industry may be racing toward self‑improving systems without proven alignment methods.
- Amodei said his unilateral move to host outside evaluators could buy researchers one to two years to harden safeguards but would likely need government action on legal rules and export controls to prevent competitors from pulling ahead.