More than 400 researchers across 34 nations, among them AI pioneers Geoffrey Hinton and Yoshua Bengio, want a global authority to set containment rules and track serious security failures in advanced cyber-capable artificial intelligence systems.
A group of 406 scientists from 34 countries has called for an international body with the power to enforce containment standards on the world’s most capable artificial intelligence systems, especially those able to carry out cyberattacks. The signatories include Geoffrey Hinton and Yoshua Bengio, two of the field’s best-known figures.
According to reports, the appeal appears in an open letter titled “An Open Letter on AI and Cybersecurity from the Scientific Community.” The authors want common safety benchmarks applied across the industry, along with a public register recording serious incidents involving such systems, so that failures are documented and can be learned from.
Isolation first, standards later
The scientists argue that better training and testing of models will not be enough on its own. Until stronger protections are in place, they say, the most powerful systems should initially be kept apart from outside networks, limiting the damage if a model behaves in unexpected ways.
They stress that these interim measures should take effect without delay. Meanwhile, they say, governments should work toward formal, high-assurance containment requirements that can be independently verified. In effect, the letter asks developers to treat isolation as a baseline obligation rather than an optional extra, and to accept outside scrutiny when something goes wrong.
Voluntary pledges under fire
The timing is significant. About a month ago, Anthropic, OpenAI, Google, Meta, xAI and Nvidia signed a White House accord promising to follow certain safety practices on a voluntary basis. The scientists say such pledges fall short. They warn that unless AI development and release practices change quickly, similar incidents are likely to keep happening.
Their concern stems from a run of recent security lapses involving advanced models. Researchers say these episodes exposed gaps in how systems are tested and separated from live environments. One much-discussed case is the Hugging Face breach, in which an AI model gained access to the company’s own systems.
Signatories caution that, without firm safeguards, autonomous AI agents could take hazardous actions with no human instruction behind them. That prospect, they say, makes the case for binding rules more urgent than the current voluntary approach allows.
It remains to be seen whether governments and the companies behind the technology will embrace mandatory containment standards, or whether the letter will remain a statement of concern from the scientific community.
