Anthropic CEO Dario Amodei is calling for a faster and more forceful regulatory response to artificial intelligence, arguing that policy is moving too slowly for a technology advancing at exceptional speed.

In a new policy essay, Amodei says AI systems have progressed from producing only limited code to generating most of the code at major AI companies in just a few years. He argues that this rapid improvement has pushed AI into a category of technology with major strategic and safety implications, while lawmakers and regulators are still operating on a far slower timeline.

Amodei frames the problem as a widening gap between what AI can do and what political institutions can absorb. While governments often move cautiously for good reason, he says the pace of AI development means the delay between identifying a risk and enacting a response can leave policymakers behind the curve. He says the first wave of AI policy focused largely on transparency and information gathering because the risks were still uncertain, but that the situation has changed.

According to Amodei, recent evidence has made the dangers more concrete, particularly in cybersecurity. He points to frontier models, including Anthropic’s own Claude Mythos Preview, as examples of systems that revealed serious cyber-related vulnerabilities. In his view, these developments show that AI models are no longer just consumer tools but technologies with national and global security consequences.

That assessment leads to his central policy recommendation. Amodei says AI regulation should start to resemble the oversight used for airplanes, drugs or other high-risk industries. He argues that frontier models should undergo mandatory technical testing and independent auditing before release, and that deployment should be restricted if they fail to meet safety standards.

Anthropic’s proposal, released alongside the essay, calls for third-party testing of models above a certain compute threshold. The testing would focus on four areas: cybersecurity, biological risks, loss of control of AI systems and automated research and development that could speed up those threats. The company also wants AI firms to maintain strong security protections for model weights, conduct regular red-teaming and penetration testing, and report serious incidents promptly.

The proposal leaves open whether evaluations would be conducted by a government agency or by private organizations authorized and inspected by the government. Amodei says either approach could work as long as there is a system with real enforcement power and safeguards against political interference.

He also acknowledges that even this framework may not be enough if AI capabilities continue to advance. In that case, he says, governments may need tougher restrictions later, especially if systems begin to resemble highly dangerous materials rather than ordinary software tools. But for now, he argues, policy should target the risks that are already emerging rather than waiting for a more dramatic crisis.

The essay also links regulation to broader economic and social questions. Amodei says powerful AI could accelerate growth while also disrupting labor markets at a scale that older technologies did not. He warns that the main challenge could become how to distribute the gains from AI, rather than how to encourage growth itself.

Anthropic says the essay is part of a broader push to shape AI governance as the technology becomes more powerful. The company says it is backing the legislative proposal and a separate framework on job displacement as early steps in a longer policy effort.