Scientists urge cyber-capable AI models be kept offline in testing

An international group of 420 scientists has signed an open letter asking developers, deployers and testers of cyber-capable AI models to disconnect them from the internet at first, according to EdTech Innovation Hub.

The letter said model escapes from restricted environments stem from both unpredictable AI behaviour and weak cybersecurity practice. It states that "both improved AI alignment, leading to more predictable model behavior, and stronger sandbox technology are required."

Signatories include Geoffrey Hinton of Toronto, Yoshua Bengio of Montréal, Stuart Russell of Berkeley and Herbert Bos of VU Amsterdam. Researchers from Oxford, Cambridge, UCL and Imperial are also among them.

The letter makes four demands. It calls for high-assurance containment to be adopted now and for models to be airgapped as an initial precaution. It also asks that permitted actions be limited for models that are connected, and that alignment and sandboxing be advanced together.

It also calls for an international body, built with governments and AI and cybersecurity experts. That body would set containment and alignment standards, including on internet access. Containment failures would be reported to it and eventually made public.

The letter sets no timetable and names no organisations or incidents.

It arrives as Anthropic has cut live internet access from its internal evaluations and the White House has ordered incident reporting, according to the report.

The letter does not say when the proposed international body would be formed, and no timetable for any of the demands has been set.