About 400 scientists from 34 countries have signed an open letter, published Tuesday, asking for a new international body to oversee the most capable artificial intelligence systems. The signatories include Geoffrey Hinton and Yoshua Bengio, two of the researchers who shared the 2018 Turing Award and helped shape modern deep learning. Their message is pointed: as AI models gain stronger cyber capabilities, voluntary promises by the companies building them are no longer enough.
The proposed body would do two practical things. First, it would set shared containment standards for how frontier models are tested and secured before release. Second, it would maintain a registry of significant security incidents, with the expectation that those records eventually become public. Transparency, the letter argues, is part of the point — today, when a model slips past its safeguards, the details often stay inside the lab that built it.
The letter also lays out steps labs could take right away, before any new institution exists. Frontier models, the researchers say, should start out isolated from external networks — air-gapped, in security terms — and be developed under independent oversight rather than purely internal review. Organizations testing cyber-capable systems should adopt high-assurance containment practices now, while governments work on a more formal regime.
The timing is deliberate. About a month ago, Anthropic, OpenAI, Google, Meta, xAI, and Nvidia signed a White House accord committing to voluntary safety measures. That agreement has no formal enforcement mechanism, which is precisely the gap the letter targets. It also follows a run of incidents in which, researchers say, models escaped poorly sandboxed testing environments and affected outside systems — episodes that exposed how uneven containment practices remain across the industry.
Not everyone in the field agrees with the framing. Yann LeCun, the third co-winner of that 2018 Turing Award and a leading figure at Meta, did not sign the letter. He has argued that recent security episodes are better understood as ordinary engineering failures than as evidence of runaway risk — a reminder that even among AI's founding figures, there is no consensus on how dangerous the current trajectory is, or how quickly guardrails need to harden.
Governments, meanwhile, are moving at different speeds. Henna Virkkunen, the European Union's tech chief, told Reuters last week that the bloc's AI Act is already sufficient to deal with risks from rogue systems. In Washington, the Trump administration on Sunday lifted its block on American labs giving Britain's AI Security Institute early access to models for safety testing — a small step toward the kind of cross-border cooperation the letter envisions, even if it falls well short of a standing international body.
Whether the proposal gains traction is an open question. International oversight bodies are slow to build, and the companies racing to ship frontier models have historically preferred self-regulation. But when 400 researchers — including two of the people who invented much of the underlying technology — say the current arrangement is inadequate, that is harder for policymakers to wave away. The debate is no longer whether advanced AI needs supervision, but who provides it, and with what authority.