Dario Amodei, the CEO of Anthropic, published a detailed policy essay this week calling on governments to stop treating AI oversight as a voluntary exercise. The core demand: give regulators the legal power to block or reverse an AI model's deployment if independent testing finds it poses unacceptable risks. That authority does not currently exist in any comprehensive form in the United States.
The proposal is modelled on how aviation works. Before a new plane type carries passengers, it goes through mandatory independent testing. Amodei wants the same logic applied to the most powerful AI models. Any system trained above a certain computational scale, or built by a company with over $500 million in annual AI revenue, would face mandatory third-party testing across four specific risk areas: cybersecurity attacks, biological weapons, loss of human control over the AI system, and automated research that could accelerate those risks further.
The cybersecurity concern is not abstract. Amodei points to Anthropic's own Claude Mythos Preview as evidence. That model was found capable of discovering serious software vulnerabilities across major operating systems at a scale that has rattled security professionals. When the CEO of the company that built the tool says it worries him, that is worth taking seriously.
On jobs, Amodei does not soften the message. He writes that AI could perform most cognitive tasks better than humans within one to two years, producing rapid economic growth alongside large-scale job displacement at the same time. His proposed tools include wage insurance for workers who are pushed into lower-paying roles, tax incentives for companies that retain staff rather than replace them, and, as a longer-term option, a universal basic income funded by taxes on AI companies and capital gains.
The proposals go well beyond what Washington is currently debating. Amodei also calls for faster drug approval processes for AI-discovered medicines, a ban on fully autonomous weapons for domestic use, and stronger controls on the export of advanced computer chips.
There is an obvious tension worth naming. Anthropic filed confidentially for an IPO at a valuation of roughly $965 billion, with a revenue run rate of $47 billion, just days before this essay appeared. The rules Amodei proposes would apply most heavily to a small number of large frontier developers, which describes Anthropic exactly. Critics have pointed out that proposing regulation you can already meet is a way to raise the barrier for competitors who cannot. That reading is not unfair.
Amodei's response to that charge, stated or implied throughout the essay, is that the risks he describes are real whether or not the timing is convenient. His own model demonstrated live cybersecurity vulnerabilities. His company has observed self-improving AI behavior multiple times in its own internal tests. He is not describing future risks: he is describing things that have already happened inside his own walls.
For professionals in any industry that relies on software, financial systems, supply chains, or anything connected to the internet, the cybersecurity argument is the most immediately relevant. If AI can now find and potentially exploit vulnerabilities in major operating systems at scale, the threat surface for every organization just expanded. That is true regardless of whether Amodei's regulation proposal ever becomes law.
The jobs argument will take longer to play out, but Amodei is not alone in raising it. The CEO of JPMorgan and analysts at Goldman Sachs and Boston Consulting Group have all flagged meaningful job displacement in the near term. Amodei is unusual in proposing specific policy responses and putting Anthropic's money behind them: the company announced a $200 million investment into research on AI's impact on society alongside the essay.
Whether governments move fast enough to matter is a separate question. The essay is a serious document, but it is still a proposal from a private company, not a law. The gap between what Amodei is asking for and what Washington is currently considering is wide.