Anthropic CEO urges mandatory audits after Claude Fable 5

Anthropic released Claude Fable 5 on June 9. The next day, CEO Dario Amodei published an essay titled ‘Policy on the AI Exponential’ and proposed mandatory third-party testing for frontier AI systems, plus government authority to block releases that fail safety audits.

Amodei proposed a certification regime modeled on the Federal Aviation Administration. Models that exceed a defined compute threshold would face independent audits in four technical areas: cybersecurity, biological threats, loss of control, and automated AI research. The proposal includes prompt reporting of safety incidents and legal protections for model weights. He described the White House executive order on AI as a limited step forward.

“Frontier AI models, like airplanes, should be required to go through technical testing and auditing, and their release should be blocked or reversed as a threat to public safety if they do not meet high standards of safety,” Amodei wrote.

Anthropic singled out cybersecurity as the most immediate risk. The company pointed to a Mythos preview model that solved 73% of expert-level cyber challenges that prior systems had not passed. The essay warned such capabilities could be used to attack financial systems and critical infrastructure. Anthropic stated that Claude Fable 5 includes safeguards to block high-risk cyber and biological requests.

The essay raised concerns about increasing model autonomy and about AI systems improving other AI. Anthropic’s internal data shows advanced models writing substantial portions of code at major AI labs, a development the paper linked to potential loss-of-control scenarios.

On economic issues, the paper proposed wage insurance, retention tax incentives for employers, and workforce training grants. If job losses prove long-lasting, Amodei suggested funding universal basic income through company taxes or higher capital gains taxes.

On civil liberties, the essay called for a ban on fully autonomous weapons in domestic law enforcement and urged Congress to close a data broker loophole that enables bulk surveillance purchases. On supply chains, it recommended a coalition of democracies to control access to advanced chips and semiconductor manufacturing equipment and cited pending US bills MATCH and OVERWATCH as early steps in coordinated export controls.

Whether Congress or regulators will adopt an FAA-style testing regime is uncertain. Amodei framed the plan as a tool for governments to block or reverse releases that fail safety audits, writing that such powers are needed because AI is progressing faster than current policy processes.