Anthropic CEO: Governments should have the authority to block or restrict the deployment of high-risk new AI models.

Anthropic CEO Dario Amodei has publicly called for stricter mandatory regulation of artificial intelligence, advocating that governments should have the authority to block the deployment of AI models posing safety risks—a stance that clearly diverges from the Trump administration’s current emphasis on voluntary compliance.

In a lengthy post on Wednesday, Amodei proposed that AI models be required to undergo third-party testing covering multiple risk dimensions, including cybersecurity threats and biological weapons.

He wrote that if a model is deemed to pose "unacceptable risks," "the government should have the authority to block or restrict its deployment." This marks one of his strongest statements to date on AI regulation.

This position stands in stark contrast to the White House’s current policy. Earlier this month, Trump signed an executive order granting the government only voluntary access to AI models, without requiring developers to obtain explicit approval prior to deployment.

Regulatory Proposal: From Transparency Toward Mandatory Constraints

In his post, Amodei drew analogies to automobiles, aircraft, and pharmaceuticals, arguing that AI, like these technologies, is an indispensable yet powerful tool in the modern economy—capable of causing mass casualties if improperly designed or operated.

"It is time to move beyond transparency toward more serious and binding AI regulation," he wrote. He also warned that global policy mechanisms are responding too slowly, while the risks and opportunities presented by AI will "compound with unexpectedly rapid speed."

Just last week, Anthropic also called for establishing a mechanism through which governments and AI developers could jointly decide when to slow down research and development in response to potential dangers. These two public statements indicate that the company is prioritizing the establishment of a regulatory framework as a strategic objective.

Mythos Model: Internal Deliberations Prompted by Capability Boundaries

Amodei cited Anthropic’s new model, Mythos, as an example in his post, describing it as "a perfect illustration of AI’s astonishing capabilities and potential risks."

Anthropic previously disclosed that Mythos has the capability to identify and exploit critical software vulnerabilities, prompting the company to restrict access to the model to only a limited number of partners. On Tuesday, Anthropic released a separate version with its cybersecurity capabilities removed, making it available to a broader user base.

These decisions reflect an inherent tension among leading AI developers between product capabilities and safety boundaries—while publicly calling for regulatory intervention, companies are already implementing their own forms of screening and restrictions akin to regulation.

Policy Divergence: Misalignment Between Industry Appeals and Government Direction

Amodei's remarks place Anthropic at odds with the current federal policy direction. The Trump administration’s executive order favors voluntary mechanisms over mandatory approvals, reflecting this administration’s overall preference for minimizing regulatory intervention in the AI sector.

Anthropic has long positioned itself as a 'responsible AI developer,' and its recent public advocacy for granting the government mandatory intervention authority signals its intent to play a more proactive role in shaping regulatory discourse.

However, given Washington’s current policy environment leaning toward regulatory leniency, there remains considerable uncertainty as to whether such appeals will translate into substantive legislation or regulatory action.

Editor/Liam