Anthropic CEO Dario Amodei is warning that governments are moving too slowly to keep pace with frontier AI systems and that voluntary safety efforts are no longer enough. In a new policy essay, he argues that the next phase of AI governance should require mandatory third-party testing for the most advanced models and give regulators the power to stop deployments that pose serious public risks.
Amodei says the gap between AI progress and public policy has become more dangerous as models have grown more capable across software development, science, finance, law, translation, and other fields. He compares the pace of AI improvement with the slower rhythms of legislation, saying that systems can transform in just a few years while governments often need years to act.
The essay marks a shift in emphasis from transparency measures, which Anthropic has supported in recent years, toward more binding rules. Amodei writes that earlier on it was difficult to know exactly how AI risks would appear, making rigid laws hard to design. That led safety advocates to focus on disclosure, incident reporting, and other steps that would improve visibility into emerging dangers without locking regulators into outdated requirements.
Now, he says, the case for stricter oversight is stronger. He points to recent evidence that frontier models can pose serious cybersecurity risks and potentially affect critical infrastructure, financial systems, and national security. He argues that these risks show AI has reached a level of strategic importance that requires a more serious regulatory response.
Anthropic is pairing the essay with a legislative proposal that would require frontier models above a certain compute threshold to be tested by qualified third parties. Those tests would focus on four risk areas: cybersecurity, biological weapons, loss of control of AI systems, and automated research and development that could make those threats worse.
Under the proposal, regulators would also be able to block or delay deployment if an independent assessment finds unacceptable risks in those categories. Amodei says such a system should be tightly limited so that it cannot be used for political favoritism or arbitrary intervention.
He also calls for stronger security standards for companies building advanced models, including protections for model weights, regular red teaming, penetration testing, and coordination with government on major threat actors. Safety incidents in the four designated risk areas would need to be reported quickly.
Amodei suggests that the overall model should resemble aviation regulation more than light-touch tech policy. In his view, frontier AI should be treated like a powerful technology that is useful to society but dangerous enough to justify technical review before release.
Alongside the essay, Anthropic said it will financially support the new legislative proposal and a separate policy framework on job displacement. The company described the two efforts as early steps in a broader policy agenda.
The essay also touches on the possibility that AI could trigger major economic disruption. Amodei argues that if systems become capable of handling most cognitive work better than humans, they could accelerate growth while also pushing labor markets into more unequal territory. He says the challenge would not just be generating growth, but making sure people share in the gains.
While the essay is focused on U.S. policy, Amodei says the ideas are relevant internationally as governments grapple with rapidly advancing AI systems and the risks that come with them. His broader message is that the policy response must become more concrete, more enforceable and faster, before the next wave of capability arrives.