Anthropic CEO Dario Amodei published an essay arguing that transparency-first AI regulation is no longer sufficient and that the US needs binding safety requirements for frontier AI models, modeled on the FAA's approach to aircraft certification. The proposal mandates third-party testing across four risk categories: cybersecurity, bioweapons, loss of AI control, and automated R&D. Companies would be required to secure model weights, conduct safety testing, and report serious incidents, with government authority to block or reverse deployments that threaten public safety. Amodei also calls for preparing for AI-driven job displacement and advances in drug development, limiting surveillance and autonomous weapons in domestic law enforcement, and strengthening democratic cooperation on critical AI technologies. The essay comes as Anthropic expands access to Claude Mythos with the launch of Mythos 5, a frontier model that can autonomously execute complex cyber attacks. Critics, including OpenAI CEO Sam Altman, have accused Anthropic of using fear-based marketing to justify concentrating control of AI technology. Amodei rejects this, stating that public concern reflects legitimate risks and democratic accountability.

Key facts

• Amodei argues transparency-only AI regulation is outdated; binding safety requirements are needed.
• Proposal mandates third-party testing for cybersecurity, bioweapons, loss of AI control, and automated R&D.
• Government would have authority to block or reverse unsafe AI model deployments.
• Anthropic launched Mythos 5, a frontier model capable of autonomous cyber attacks.
• Critics accuse Anthropic of using fear-based marketing to justify AI control consolidation.

KeyAudit data perspective