Summary

  • Dario Amodei asserts that the period of prioritizing transparency in AI regulation has ended, advocating for FAA-like testing standards for advanced AI models.
  • The essay emphasizes mandatory third-party evaluations in four key risk areas: cybersecurity, bioweapons, AI control loss, and automated research and development.
  • Alongside the essay, Anthropic plans to introduce a legislative proposal on frontier model testing and policies regarding job displacement.

On Wednesday, Dario Amodei, the CEO of Anthropic, emphasized that AI regulation can no longer be viewed merely as a subject for study, urging the U.S. to implement mandatory safety measures for its most robust AI systems.

In his essay titled "Policy on the AI Exponential," Amodei argues that mere transparency is insufficient and insists on the need for enforceable regulations for cutting-edge AI technologies.

“The pace of AI advancement is astounding—within just four years, AI systems have evolved from struggling to produce coherent code to generating most of the programming used by leading AI firms,” Amodei stated.

This essay appears as Anthropic expands the availability of its Claude Mythos through the introduction of Mythos 5 on Tuesday, a specialized frontier AI model designed for cybersecurity entities and government collaborators. Research conducted by the UK’s AI Security Institute has revealed that it can autonomously conduct sophisticated cyber attacks.

Amodei's regulatory framework proposal draws inspiration from the Federal Aviation Administration's structure.

“Similar to airplanes, frontier AI systems should undergo rigorous technical testing and audits, and their deployment should be restricted or reversed if they pose a public safety risk,” he explained. “I appreciate the Trump administration’s Executive Order for moving toward enhanced government intervention in AI, though Anthropic’s recommendations call for even more robust actions.”

Today I'm releasing a new essay, Policy on the AI Exponential. AI is advancing at a rapid rate—much quicker than the policy framework was designed to manage. The essay outlines the current state of the technology and the necessary actions to bridge the gap: https://t.co/Lh6PWae178

— Dario Amodei (@DarioAmodei) June 10, 2026

Under Amodei’s plan, a regulatory system would necessitate third-party assessments of advanced AI models, empower the government to prevent unsafe implementations, and require companies to secure model weights, conduct safety evaluations, and report significant incidents. He urges governments to brace for job losses driven by AI advancements, advancements in drug creation, and to limit surveillance and autonomous weaponry in domestic policing while enhancing collaboration among democratic nations on critical AI technologies.

“Firstly, persistent job displacement is both undesirable and perilous, and we should strive to mitigate or avert it, rather than allow it to occur,” he remarked, referencing previous warnings he has made regarding job losses. “Secondly, any measures taken in response to AI-related job displacement must address the economic needs of all individuals, alongside the necessity for people to experience meaning, purpose, and agency.”

The essay coincides with the launch of Claude Fable 5, a public-facing iteration of Claude Mythos 5, which directs certain inquiries related to cybersecurity, biology, chemistry, and AI development to the less advanced Claude Opus 4.8 to prevent misuse. This launch has faced backlash from developers and researchers due to Fable's increased token consumption, compulsory 30-day data retention policies, and safeguards that may diminish the model's functionality without prior notice to users.

Amodei's push for regulatory change in AI development comes at a time when Anthropic is gearing up for an IPO. Earlier this month, the firm submitted its initial public offering paperwork to the U.S. Securities and Exchange Commission after securing a reported $65 billion in a Series H funding round at an estimated valuation of $965 billion.

While Amodei positions the issue as a contest between technological advancement and public policy, some critics have expressed skepticism regarding whether calls for stricter AI regulation genuinely benefit the public. In April, OpenAI CEO Sam Altman accused Anthropic of employing "fear-based marketing" to promote Claude Mythos, suggesting that fears surrounding advanced AI could be leveraged to justify consolidating control of the technology among a select few companies.

“There are many ways to rationalize that, and some concerns are legitimate, such as real safety issues,” Altman noted. “However, if the goal is to assert control over AI exclusively for ourselves, under the guise of being the trustworthy entities, I believe fear-based marketing is an effective method to validate that.”

Amodei dismissed the notion that worries about advanced AI are merely a public relations challenge, asserting that the apprehensions surrounding the technology stem from genuine issues that need to be addressed.

“People are concerned about AI because they accurately recognize that its risks are substantial, not due to a lack of optimism from AI executives,” he stated, referencing the character Pangloss from Voltaire's Candide, known for his unwavering optimism regardless of circumstances.

“I see it as my responsibility as a leader in AI to maintain transparency regarding these risks, and the public's concerns in response to that transparency are a manifestation of democratic accountability functioning as intended,” he added.