Anthropic CEO Dario Amodei is calling for mandatory third-party testing of frontier artificial intelligence models, arguing that voluntary safeguards can't keep pace with rapidly advancing capabilities. In an essay published June 10 — one day after the company released Claude Fable 5 — Amodei laid out a regulatory framework modeled on the Federal Aviation Administration, complete with pre-deployment audits and the power to block systems that fail safety checks.
From Transparency to Hard Rules
Anthropic has historically backed disclosure-based laws such as California's SB 53, the federal RAISE Act, and SB 315. But Amodei now says that approach is no longer enough. “Transparency alone is insufficient,” he wrote, pointing to the growing gap between industry safeguards and real-world risk. Voluntary commitments, he argued, cannot substitute for binding rules that apply to every major developer.
The shift comes after Anthropic's own Claude Mythos Preview solved 73% of expert-level cybersecurity challenges that no AI had previously cracked. The feat, Amodei noted, shows that frontier models are already pushing into territory where mistakes could have serious consequences.
An FAA-Style System for AI
Under Amodei's proposal, any model trained above a certain compute threshold would face mandatory audits in four areas: cybersecurity, biological weapons, loss of control, and automated AI research. Developers would also be required to report safety incidents promptly and to protect model weights from unauthorized access.
Governments, he said, should have the authority to block deployment of systems that fail those audits. The regime borrows from aviation safety: just as an aircraft can't fly without certification, a powerful AI model shouldn't hit the market without a green light from independent testers.
Claude Fable 5, the latest model, already includes safeguards that block high-risk requests related to cyberattacks and bioweapons. But Amodei warned that voluntary guardrails won't stop a competitor from skipping them. “We need binding rules across the industry,” he said.
Economic and Geopolitical Proposals
The essay didn't stop at safety audits. Amodei also called for wage insurance, retention tax incentives, workforce training grants, and universal basic income financed through company or capital gains taxes — a recognition that AI could disrupt jobs and critical infrastructure.
On the geopolitical front, he urged Congress to close the data broker loophole that allows bulk surveillance purchases, and to ban fully autonomous weapons from domestic law enforcement. He also pushed for a coalition of democracies to control chips and semiconductor manufacturing equipment, citing pending U.S. bills MATCH and OVERWATCH.
Amodei rejected the idea that public fear of AI is a marketing problem. “The concern is accurate,” he wrote, dismissing efforts to frame anxiety as a communication challenge.
What Lawmakers Are Watching
Congress has yet to take up a comprehensive AI bill, but the MATCH and OVERWATCH bills remain active. Whether lawmakers will embrace mandatory pre-deployment audits — let alone the broader set of economic and national security measures Amodei advocates — is an open question. For now, the industry is watching to see if anyone blinks first.
