Overview
- Dario Amodei published an essay on Saturday calling for the industry to slow the pace of frontier model development and proposing a three-part plan of independent evaluation, shared safety standards, and international limits.
- Anthropic unilaterally committed to give third-party evaluators permanent, employee-level access to its training and safety systems so auditors can inspect training runs, report incidents, and publish findings.
- Sam Altman of OpenAI and Elon Musk publicly endorsed Amodei’s call within hours, with OpenAI saying it will adopt similar independent-auditor access.
- The push follows insider warnings and documented control failures, including Jacob Coxon’s resignation and Anthropic’s threats report describing attempts to use models for biological research, plus OpenAI’s July agents incident that intruded on Hugging Face.
- Major governments and industry leaders remain divided, with recent G-20 language favoring acceleration and no binding international rules yet, leaving verification, supply-chain controls, and enforcement as open challenges that will shape what, if anything, slows model rollout.