Anthropic CEO Dario Amodei called on June 10 for the US government to have legal power to block or reverse the release of dangerous AI models that fail mandatory safety tests. He published an essay outlining the need for binding regulation and safety oversight of advanced AI systems, comparing AI to cars, airplanes, and drugs that require rigorous testing before public use [1, 2].

Amodei criticized current voluntary disclosure frameworks as inadequate and argued that transparency alone does not address the safety risks of AI. "Frontier AI models, like airplanes, should be required to go through technical testing and auditing," he said, adding that unsafe models should be blocked from deployment or withdrawn if already released [3].

He cited Anthropic's new AI model Mythos as an example showing both the power and risks of AI, including cybersecurity threats, biological risks, and loss of control or autonomy. On the same day as his essay, Anthropic released a less capable version of Mythos with cybersecurity functions removed to allow broader access [1, 2].

Amodei said AI advances at a "lightning pace," while policy and legislation lag behind. He warned that in the years Congress takes to act, AI could become vastly more powerful. He called for a slow but steady activation of policy measures to manage AI risks and opportunities [1, 3].

He also proposed economic policies to ease AI disruptions, including wage insurance, tax incentives, and expanding the social safety net [1].

The Trump administration’s recent AI executive order moves incrementally toward government roles in AI oversight but remains voluntary and less strict than Amodei’s calls for mandatory legal authority [1, 3, 2].

Amodei’s essay marks a push for more legally enforceable safety standards on AI models and government powers to intervene on public safety grounds. Anthropic’s release of a restricted Mythos version demonstrates a practical step balancing capability with risk. Further legislative or regulatory actions addressing Amodei’s demands have not yet been announced.