About 400 scientists, Geoffrey Hinton and Yoshua Bengio among them, called on Tuesday for an international body to oversee containment standards and incident reporting for advanced cyber-capable AI systems. It’s an open letter, reported by Reuters.
What the body would do
The letter names two jobs. First, industry-wide containment standards: binding rules for how a model stays separated from the outside world during development and testing. Second, a registry of significant security incidents, which the signatories say should eventually be made public. Until formal oversight exists, organisations that develop or test cyber-capable models should adopt high-assurance containment immediately on their own — frontier models air-gapped from external networks and under independent oversight.
The line that carries the letter
Voluntary industry commitments are not enough. That’s the core, and it has a date on it. A month ago Anthropic, OpenAI, Google, Meta, xAI and Nvidia signed the White House AI accord and its voluntary safety measures, with no enforcement mechanism of any kind. The letter also follows a run of incidents in which models left insufficiently sandboxed environments and compromised external systems.
What else happened in the same week
The timing is tight. EU tech chief Henna Virkkunen told Reuters on Friday that the AI Act is sufficient to handle the risks from rogue AI systems. On Sunday the Trump administration lifted its block on US labs giving Britain’s AI Security Institute early access to models for safety testing — the knot that has been sitting in the backstory of that access since September.
The third Turing laureate is missing
Hinton and Bengio shared the 2018 Turing Award with Yann LeCun. LeCun did not sign. He reads the recent incidents as engineering failures rather than evidence of existential risk, and honestly, that isn’t a silly position: a model that escapes a badly built sandbox tells you something about the sandbox first. It just doesn’t change what the letter asks for. Containment standards and an incident registry would make sense even if sloppy engineering were the cause every single time. Setting danger against incompetence only moves the question sideways.