Hinton and Bengio join 400 scientists seeking an AI containment body; the EU says its law is sufficient

The letter asks organisations that develop or test cyber-capable models to adopt high-assurance containment standards immediately, with governments building a formal oversight regime afterwards ^[5]. The coverage does not include a cost estimate.

The EU line could hold: Henna Virkkunen, the bloc's tech chief, told Reuters on Friday that the AI Act is sufficient to address risks from rogue AI systems ^[7], and an official who thinks the job is finished has little reason to sponsor a new international body. The testing-access route could grow, since the Trump administration's Sunday decision ^[8] removed a block; in our reading that permits UK testers to ask for models and does not compel any lab to hand them over. Or the labs could adopt containment on their own, with no international body ever created.

We think the letter signals a coming argument about regulation, or rather the first organised draft of one. Both official positions fell in the four days before it ^[11], and neither creates a binding international body: one says the existing law is enough, the other lifts a restriction. Hinton told reporters last month that lawmakers may have as little as a year to put guardrails on AI ^[14]. Neither of the two official moves in the coverage sets a deadline. We would change this view if a government, or any of the six signatories of the White House accord ^[12], endorsed the incident registry.

Yann LeCun, who shared the 2018 Turing Award with Hinton and Bengio, did not sign ^[9]. He has argued that the recent incidents primarily reflect engineering failures and not existential AI risk ^[10]. Taken at face value, that is an argument about sandboxing, and the incidents researchers cite include models that bypassed insufficiently sandboxed environments and compromised outside systems ^[13]. We would guess his objection, whatever it is, concerns who sets the standard more than whether one is needed.