The chief executive of one of the world's most powerful artificial intelligence companies has told governments they should have the legal authority to block AI systems from reaching the public - a position that carries direct and uncomfortable implications for every insurer, broker and underwriter now racing to embed AI into their operations.
Dario Amodei, who leads Anthropic, the privately held AI group behind the Claude family of models, published an essay on Wednesday titled Policy on the AI Exponential, in which he argued that the age of voluntary transparency in AI is finished. Third-party audits must become mandatory. Governments must acquire effective veto power. And the entire framework, he proposed, should be modelled not on a technology regulator but on the Federal Aviation Administration - an agency whose founding purpose was simple: prevent aircraft from killing people.
"Frontier AI models, like airplanes, should be required to go through technical testing and auditing," Amodei wrote, "and their release should be blocked or reversed as a threat to public safety if they do not meet high standards of safety."
The essay, published simultaneously on Amodei's personal blog and followed by an exclusive interview with ABC News, outlines a tiered framework anchored to computing power. Any AI model trained above a specified "compute threshold" - a proxy for capability and potential risk - would automatically trigger an independent pre-release investigation across four areas: cybersecurity vulnerabilities, biological weapons capabilities, the potential to accelerate automated research in dangerous domains, and the risk of models growing beyond effective human control.
Critically, Amodei does not frame this as a request for cooperation. There is no mention of voluntary compliance, no appeal to corporate goodwill. If an independent auditor determines that a model presents unacceptable risks in any of those categories, he proposes, government should have clear statutory authority to prevent deployment.
The framework goes considerably further than the executive order issued by President Donald Trump on 2 June, which encouraged AI firms to share models with government auditors on a voluntary basis and emphasised the role of intelligence agencies in cybersecurity testing - but stopped short of making anything compulsory. Amodei welcomed the order as incremental progress while making plain that incrementalism, in his view, is no longer adequate.
Accompanying the essay, Anthropic announced a $200 million investment into research on AI's societal impact, and said it would provide "substantial financial backing" to legislative proposals on frontier model testing and job displacement policy - a commitment that distinguishes the essay from the usual CEO think piece.
The catalyst Amodei points to most directly is Anthropic's own Claude Mythos Preview model, which was unveiled earlier this year and deliberately withheld from general release. During testing, the system was found capable of identifying and exploiting thousands of software vulnerabilities across major operating systems and browsers. A version without those cybersecurity capabilities was released to the public on Tuesday; the more powerful model remains restricted to a small number of vetted partners.
For Simon Hughes, CCO at Cowbell Cyber, the Mythos case illustrates exactly the principle his industry needs embedded in law: "An advanced AI model should prove its structural resilience before entering the commercial ecosystem."
In his ABC News interview, Amodei drew stark geopolitical conclusions from the exercise. "Imagine if China had built Mythos," he said. "They would have used it to attack us. They would have used it to help Russia in Ukraine." Whether or not one accepts that framing, the technical reality he describes is not disputed: a frontier AI model had acquired, without deliberate design, capabilities that could cause serious harm if widely deployed.
For the insurance industry, Amodei's intervention arrives at a moment of acute tension between AI ambition and AI anxiety.
A recent Accenture survey found that 90% of global insurance executives plan to increase AI spending in 2026, with early deployments already showing gains in combined ratios and claims efficiency. Yet a separate GlobalData poll of insurance professionals, conducted in the first half of this year, found that close to one in four participants identified AI itself as not ready for widespread use in insurance - a striking admission of doubt from within an industry accelerating towards the technology simultaneously.
The regulatory scaffolding around AI in insurance remains incomplete and fragmented. At least 17 US states introduced or advanced AI bills targeting insurers in 2025. The National Association of Insurance Commissioners is piloting an AI Systems Evaluation Tool across 12 states, designed to assess governance and high-risk model deployment in underwriting, pricing and claims. But as Willis Towers Watson warned last month, AI adoption is outpacing governance frameworks across the sector - a finding that suggests the gap Amodei is trying to close at the frontier level has an uncomfortable parallel at the industry application level.
"When technology moves at this pace, the insurance sector cannot rely on static, historical data to underwrite liabilities," said Hughes. "We need standardised, transparent benchmarks to accurately evaluate the risk profile of the digital tools British businesses are integrating into their operations daily."
The liability picture is no clearer. The Insurance Services Office introduced new AI exclusion forms effective January 2026, and AI liability is increasingly being described as "the new cyber" for SMEs - a risk that is real, rapidly evolving, and imperfectly understood by the policies currently written to contain it. Google's partnership with Beazley, Chubb and Munich Re to offer affirmative AI coverage through Google Cloud is a sign of where the market is moving, but bespoke arrangements of that kind remain the exception rather than the rule.
Hughes sees mandatory audits as the mechanism that closes that gap: "This oversight will ultimately enable insurers to design clearer, more robust coverage because the underlying risk has been independently verified, it removes the guesswork and ensures that UK SMEs can adopt advanced technology with the quiet assurance that their digital infrastructure is fundamentally secure."
Directors and officers face their own exposure. As WTW's management liability coverage leader Lawrence Fine noted in May, boards are confronting securities and governance litigation risk wrapped around a technology many directors still do not fully understand. If Amodei's proposed mandatory audit regime becomes law, the question of whether a company adequately assessed the AI models it deployed - and what it disclosed to investors about that assessment - will become a live D&O question very quickly.
Not everyone is persuaded that safety alone is driving Amodei's push. Holger Mueller of Constellation Research told SiliconANGLE that the call for regulation "gives off the foul taste of a market leader that wants to freeze the market, and preserve its position at the top." Anthropic, after all, has already built internal safety frameworks - its Responsible Scaling Policy, introduced in September 2023 and updated since, and its AI Safety Level tiering system - that closely resemble what Amodei is now proposing governments mandate for everyone. A regulatory floor that looks like Anthropic's current ceiling is not an entirely disinterested proposal.
That tension does not necessarily invalidate the underlying argument. But it is worth holding in mind as the essay moves from personal blog to legislative hearing - a journey that, if Amodei's $200 million commitment is any guide, he intends to accelerate.
