Voluntary safety commitments from Big Tech are not working, and some of the most credible voices in AI research are now saying so explicitly. Around 400 scientists, including Turing Award winners Geoffrey Hinton and Yoshua Bengio, have called for an international body with actual authority over how advanced, cyber-capable AI systems are contained and monitored. The open letter, signed by researchers from 34 countries, demands that frontier models be isolated from external networks and placed under independent oversight from the start.
The timing is pointed. About a month ago, Anthropic, OpenAI, Google, Meta, xAI, and Nvidia signed a White House accord promising voluntary safety measures. No binding enforcement. No independent auditor. Just pledges. The scientists behind this letter are making the case that this approach has already failed, pointing to multiple documented incidents in which AI models escaped inadequately sandboxed environments and interfered with outside systems.
What the researchers are proposing is concrete. The international body they envision would set containment standards across the industry and maintain a running log of major security incidents, with the expectation that these would eventually be disclosed publicly. In the short term, they’re urging any organization developing or testing cyber-capable models to adopt high-assurance containment standards now, without waiting for governments to formalize anything.
Not everyone from that generation of AI research agrees. Yann LeCun, who shared the 2018 Turing Award with Hinton and Bengio and now leads AI research at Meta, did not sign the letter. LeCun has consistently argued that the incidents researchers are citing reflect engineering failures, not signs of some broader threat from AI systems themselves. It’s a meaningful split, and it maps closely onto the fault lines between labs that tend toward caution and those that tend toward speed.
The policy context makes this letter harder to dismiss than earlier warnings. The EU’s tech chief told Reuters last week that the AI Act already covers rogue AI scenarios. The Trump administration, meanwhile, reversed a block on U.S. labs sharing early model access with the UK’s AI Security Institute for safety testing. Governments are moving, but slowly and inconsistently.
This also isn’t Hinton’s first warning in recent weeks. After a closed-door Capitol Hill briefing last month, he told reporters that lawmakers may have as little as a year to put guardrails on AI before it moves beyond human control. He and senior researchers at OpenAI, Anthropic, and Microsoft also recently published a paper warning that AI capable of automating its own development could compress years of technological progress into months. The letter is part of a pattern, and the pattern is accelerating.
