Anthropic chief Dario Amodei has published "Policy on the AI Exponential," arguing the moment for voluntary transparency has passed and frontier models now need binding regulation modelled on the Federal Aviation Administration.
His core proposal: any model above a compute threshold should face mandatory third-party testing for four risks (cybersecurity, biological weapons, loss of control, and automated R&D that could accelerate the others), with the government able to block or reverse a deployment if an assessment finds unacceptable risk. He points to Claude Mythos Preview, which he says "scrambled the global cybersecurity landscape," as proof the danger is now concrete rather than hypothetical. Anthropic is releasing a draft frontier-testing bill and a job-displacement framework alongside the essay, and says it will put real money behind both.
Coming from a frontier lab, a call for the state to hold an off switch on its own products is striking. The catch Amodei concedes from Anthropic's own voluntary frameworks: write the rules too early and 95 per cent of compliance effort can land on risks that never materialise.