Dario Amodei wants the government to ground unsafe AI models the same way the FAA grounds unsafe planes. The Anthropic CEO published a blog post on Thursday calling for mandatory third-party audits of frontier AI systems, with governments empowered to block deployment if an independent auditor flags unacceptable risks.

It is, by a meaningful margin, the most aggressive regulatory framework any major AI company leader has publicly endorsed. And it lands just days after President Trump signed an executive order on June 2 that gave intelligence agencies an enhanced role in AI model testing, but stopped well short of what Amodei is now proposing.

What Amodei is actually asking for

The proposal centers on a compute threshold. Any AI model that exceeds a specified level of training compute would trigger mandatory independent audits before it could be released to the public.

Amodei’s essay identifies four risk categories that auditors would evaluate: cybersecurity vulnerabilities, biological weapons capabilities, the potential for AI systems to slip beyond human control, and the ability to accelerate automated research and development in dangerous domains.

If an independent third-party auditor determines the model poses unacceptable risks in any of those areas, governments would have the authority to block or deter its deployment. Not suggest. Not recommend. Block.

How this compares to Trump’s executive order

Trump’s June 2 executive order took a notably softer approach. It encouraged AI companies to voluntarily share their models with the government one month before public release and reinforced the intelligence community’s involvement in cybersecurity testing.

Amodei’s proposal goes considerably further in several ways. The word “voluntary” is conspicuously absent. Where the executive order asks for cooperation, Amodei’s framework demands compliance. Where the executive order leans on existing intelligence agencies, Amodei wants independent auditors with the power to issue what amounts to a veto.

Anthropic has been building toward this position for a while. The company first released its Responsible Scaling Policy in September 2023, then updated it to version 3.0 in February 2026. That revision introduced AI Safety Levels, or ASLs, a tiered framework for managing risk as model capabilities increase. The company has also publicly supported state-level AI legislation, including California’s SB 53.

Why crypto and decentralized AI should be paying attention

If Amodei’s framework gets any traction in Washington, the implications extend far beyond Anthropic and its competitors in the centralized AI lab space. Any regulation built around compute thresholds and mandatory audits creates a compliance moat. Large centralized labs like Anthropic, OpenAI, and Google DeepMind have the resources and legal teams to navigate that kind of regime. Smaller players and open-source projects might not.

That dynamic could accelerate interest in decentralized AI protocols, particularly those built with on-chain governance and verifiable compute. If the regulatory question becomes “can you prove your model was audited and deemed safe,” then systems designed around transparency and cryptographic verification have a natural advantage.

The flip side is equally important to consider. If compute thresholds become a legal trigger point, projects that train large models without the infrastructure for independent audits could find themselves on the wrong side of regulators. Decentralization is not automatically a shield against compliance requirements, and protocols that can’t demonstrate safety testing may face the same scrutiny as their centralized counterparts.