The CEO of Anthropic just published a 5,000-word policy essay using his own company’s model as the reason the government needs mandatory AI testing. That’s not humility. That’s a company trying to get ahead of regulation it knows is coming.
Dario Amodei dropped “Policy on the AI Exponential” today on his personal site, and it’s the most comprehensive public policy proposal Anthropic has ever put out. The announcement tweet already has 410K views. He’s not whispering this.
Today I’m publishing a new essay, Policy on the AI Exponential. AI is progressing extremely fast—much faster than the policy process was built to handle. The essay lays out where I think the technology is now, and the action needed to close the gap:
— Dario Amodei (@DarioAmodei) June 10, 2026
The hook: Anthropic’s own model is the case study
Here’s the part that grabbed me. Amodei cites Claude Mythos Preview, Anthropic’s own frontier model, as the breaking point. The essay says Mythos Preview “scrambled the global cybersecurity landscape.” His own red team found the model poses “very real risks” to the financial sector, critical infrastructure, and national security.
The CEO of the company whose model caused the disruption is now using that disruption to argue for mandatory third-party testing. That’s not a coincidence. That’s a company looking at the trajectory and deciding it’s better to shape the rules than to be shaped by them.
The five proposals
Amodei lays out five policy areas. Here’s the tight version.
1. Regulation: FAA for AI. He wants mandatory third-party testing for frontier models above a compute threshold. The tests cover four areas: cybersecurity, biological weapons, AI autonomy, and automated R&D that could accelerate those other risks. The government should be able to block or reverse deployment if a model doesn’t pass. He’s proposing a regulatory body similar to the FAA or a “regulatory markets” approach with authorized private evaluators.
2. Macroeconomics: UBI and wage insurance. Amodei thinks AI could create a “hypergrowth, hyper-inequality” economy. He’s proposing wage insurance, retention tax incentives, workforce training grants, and potentially universal basic income. He’s clear: “enduring job displacement is undesirable and dangerous, and we should do everything we can to minimize or prevent it.”
3. Bio-acceleration: Reform the FDA. This one’s interesting. He’s not worried about AI regulation slowing things down for downstream applications. He’s worried about the opposite: existing regulations will bottle up AI’s benefits. He wants the FDA and EMA to start developing standards now for AI-based drug development tools so they can be adopted quickly once they work.
4. Civil liberties: Ban autonomous weapons. He wants a ban on domestic use of fully autonomous weapons, closing the data broker loophole that lets bulk data collection happen without a warrant, and a right to AI legal defense for anyone facing government action.
5. Geopolitics: Democratic AI coalition. This is the big one. Amodei compares AI’s strategic importance to nuclear weapons. He wants democracies to form a coalition that shares chips and semiconductor equipment internally while denying them to adversaries. Mutual defense. Export controls. The whole package.
What this means for AI builders
If the FAA-style proposal gains traction, frontier model developers face mandatory pre-release testing. The compute threshold for mandatory testing will define who’s “frontier” and who’s not. That threshold is the single most important number in this entire essay, and Amodei doesn’t specify it.
Open-source models may be affected depending on where that line falls. If it’s set too low, it catches everyone. If it’s set too high, it catches no one. The fight over that number will be the real policy battle.
For operators and builders, this means more compliance overhead in the near term. But it also means more clarity. Right now, every company is making up its own safety standards. A regulatory framework, even an imperfect one, creates a floor.
The self-interest angle
I’m not going to pretend Anthropic doesn’t benefit from this. Regulation that raises barriers to entry favors companies that can afford to pass the tests. The company that’s already doing third-party evaluations gets an advantage over the company that isn’t.
Amodei is explicit: “We view these as first steps to signal our seriousness.” He’s not hiding the play. But the proposals are also genuinely substantive. The Mythos Preview cybersecurity findings are real. The scaling laws are real. The job displacement risk is real. You can acknowledge the self-interest and still take the proposals seriously.
What to watch next
The legislative proposal and economic framework are live as PDFs on Anthropic’s site. And there’s already been a clash with the Trump administration over defense department ties earlier this year. Here’s what matters going forward:
- Congressional co-sponsors. Without them, this is just a white paper. With them, it’s a bill.
- Competitor response. OpenAI, Google, and Meta haven’t weighed in yet. If they oppose the framework, it stalls. If they endorse it, it moves.
- The compute threshold. This is where the lobbying will happen. Every company will try to push the line in a direction that favors their architecture.
- International uptake. The democratic coalition idea is ambitious. Whether anyone outside DC listens will determine if it’s a real strategy or just a wish list.
Anthropic just drew a line in the sand. Whether anyone picks it up is the question.
Based on published reporting. Anthropic’s “Policy on the AI Exponential” essay and companion policy frameworks were published June 10, 2026.
