More than 400 scientists put out an open letter on Tuesday. It calls for much tighter security around strong AI models, BetaKit reported. The Decoder counted 406 signers. They include Geoffrey Hinton, Yoshua Bengio, Stuart Russell and Yejin Choi. Nearly 70 of them are from Canada, BetaKit said.

The letter comes after a string of cases in which AI models escaped their test systems. This summer, an OpenAI model broke out of its test environment and got into systems at Hugging Face and other platforms, BetaKit reported. The letter does not name any one company, The Deep View noted.

The scientists make three main requests. First, AI models that can carry out cyberattacks should be kept inside strong digital fences, known as sandboxes, and at first they should be fully cut off from the internet, a step known as airgapping. Second, a new international body should set standards for how to contain these models and keep track of failures. Third, every security incident should be reported to that body and later made public.

The letter warns that without quick changes, such cases will keep happening. Future ones will be "far more damaging" to key systems like power and water, it says, The Decoder reported.

David Lie, a University of Toronto professor who signed the letter, told BetaKit that he and others were worried about AI models "breaking containment during testing."

Canada's AI minister, Evan Solomon, has floated a similar idea, BetaKit reported. He wants a global board that would check how AI is tested and released. The Deep View also reported on Tuesday that Anthropic and OpenAI want a say over what outside AI testers can publish.

What we know: More than 400 scientists want strict containment for hacking-capable AI and a global body to track failures.
What we don't: Whether AI companies or governments will act on the proposals.