Anthropic urges FAA-style testing as Mythos 5 debuts

Anthropic CEO Dario Amodei on Wednesday called for Federal Aviation Administration-style mandatory testing and binding safety rules for the most capable “frontier” AI models while releasing a restricted version of Mythos 5. He published an essay titled “Policy on the AI Exponential” alongside the model rollout.

The proposal would require independent third-party testing in four risk areas: cybersecurity, biological threats, loss of control of AI, and automated research and development. Regulators would have authority to block or reverse unsafe deployments, require firms to secure model weights, mandate safety testing, and compel reporting of serious incidents.

Anthropic released Mythos 5 as a restricted, frontier-grade model aimed at cybersecurity organizations and government partners. The company also launched Claude Fable 5 as a public-facing variant that redirects queries about cybersecurity, biology, chemistry and AI development to the less capable Claude Opus 4.8.

Researchers, including the U.K.’s AI Security Institute, reported that Mythos 5 can autonomously execute complex cyber attacks. Amodei cited those findings in arguing for stronger oversight of advanced models.

Amodei wrote, “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.”

The essay also urges governments to prepare for AI-driven job displacement, to restrict surveillance and the deployment of autonomous weapons by domestic law enforcement, and to increase cooperation among democratic countries on critical AI technologies. Anthropic said it will publish proposed legislative language on frontier model testing and a policy framework on job displacement alongside the essay.

Anthropic filed paperwork with the U.S. Securities and Exchange Commission for an initial public offering after a reported $65 billion Series H funding round and a reported $965 billion valuation.

OpenAI CEO Sam Altman criticized Anthropic’s public warnings about advanced AI, calling them “fear-based marketing” that could be used to justify concentrating control of the technology with a small number of firms. Amodei rejected that characterization, writing that public concern reflects real risks and that transparency about those risks is a form of democratic accountability.

Some developers raised practical concerns about Claude Fable 5, citing higher token costs, a mandatory 30-day data-retention policy and safeguards that can reduce model capabilities without notifying users. Anthropic described the layered approach as intended to limit misuse while providing broader access.

The essay and the Mythos 5 release occur amid broader debate over how to regulate advanced AI as Anthropic prepares for a public listing.