Claude Mythos Preview scrambled the global cybersecurity landscape, and Dario Amodei is using that fact to make a blunt case: AI is now a strategic weapon, and policy is still moving at hobbit speed.
In a new essay from June 2026, the Anthropic CEO lays out the numbers that have been accumulating for years. Four years ago, AI models could barely write a coherent line of code. Today they write most of the code at major AI companies. Scaling laws—now backed by over a decade of empirical evidence—predict a continuing exponential increase in general cognitive capabilities with more compute. Amodei warns that if those laws hold for just another year or two, we get what he calls Powerful AI: "a country of geniuses in a datacenter."
The Hobbit-and-Treebeard Problem
Congress takes years to act. AI goes from amusing toy to world-altering capability in months. Amodei borrows from Tolkien: policy is Treebeard, and we're all Hobbits trying to wake him up before the forest burns.
For years, safety advocates (including Anthropic) focused on preserving optionality—transparency legislation, chip export controls, labor data collection. Admirable, but insufficient. The evidence of AI's radical power has now become undeniable. Mythos Preview proved frontier models pose concrete cyber risks: disruption of the financial sector, critical infrastructure, and national security. That's not a hypothetical; it's a live demonstration.
Amodei predicts biological risks will follow soon, and serious AI autonomy risks not far behind. He's not hedging—he says "we now, globally and collectively, need to activate a slow and rickety policy apparatus."
Five Policy Areas That Need Rewriting
The essay sketches five domains that require a complete rethink: regulation and public safety, macroeconomics and tax policy, scientific innovation, the state-society balance of power, and geopolitics. Amodei writes from a US perspective but insists most recommendations apply worldwide.
Crucially, this isn't just a think piece. Anthropic is releasing a legislative proposal on frontier model testing and a policy framework for job displacement—with "substantial financial backing" behind both. They view these as first steps to signal seriousness.
Amodei spells out the core tension: regulation can reduce harm but also blunt innovation. The Collingridge dilemma—impacts of a technology are hard to anticipate until it's too late—loomed large for AI in 2023-2024. That window has now closed. Mythos Preview was the proof of concept that forces the question: will Treebeard act before the forest is gone, or after?
Source: Policy on the AI Exponential
Domain: darioamodei.com
Comments load interactively on the live page.