Anthropic wants the government to be able to block AI models. It already can.
TL;DR: Anthropic's pitch was a safety veto on frontier models. Days later the government pulled Fable 5 and Mythos 5, for export controls and national security. Oddly, that early shutdown is useful: it shows how the government acts, and should shape the rules before a real emergency.
On Wednesday, Dario Amodei asked the government to hold the power to block Anthropic's own model releases.
On Friday, the government showed it does not need to wait for that. Citing export controls and a suspected jailbreak, it ordered Anthropic to suspend Fable 5 and Mythos 5 for foreign nationals, including its own foreign-national staff.
So what is this ask for regulation about?
Highlights:
- The main ask is for the government to be able to veto unsafe releases. Anthropic wants third-party testing in four areas: cyber, bio, loss of control, automated R&D.
- Anthropic's example of the good baseline is how they handled Mythos Preview, managing and disclosing the risk.
- Dario concedes AI may cause permanent job loss, possibly intrinsic to the technology, with universal income or capital accounts as the endgame.
- For the science AI accelerates, he wants the reverse: less regulation, not more. He warns AI will overload the FDA, where drug approval already takes 7 to 8 years, so agencies should start accepting AI evidence like simulated toxicology and synthetic control arms to speed it up.
- Government can't be fully trusted with AI. He wants autonomous weapons a court can switch off, a ban on using them inside the US, and an end to agencies buying citizens' data instead of getting a warrant.
- He treats AI as nuclear-grade geopolitics, not trade policy. A "country of geniuses in a datacenter" becomes the dominant source of military and economic power. His answer is a democratic coalition that shares chips and chipmaking equipment among members and denies them to China.
My take:
- The government may already have this power. Export controls alone were enough to force Anthropic to shut Fable 5 and Mythos 5 down for everyone, since it could not separate US from non-US users.
- Counterintuitively, this shutdown is actually good. It gives an early example of what happens when the government steps in and calls the shots, maybe without all the information it needs. Hopefully this incident shapes how the rules get rewritten.
- Now my take on the theory that the policy is really meant to slow down Anthropic's US competitors. It seems unlikely, because the major labs, Anthropic's competitors, have already built testing, guardrails, and risk management similar to Anthropic's. They just may not talk about it much in public. The chip controls, though, are a real mechanism that will keep countries outside the coalition behind. I am not sure how that plays out once frontier model capabilities plateau.