AI cyberattack models must be trained in isolation, hundreds of researchers jointly call

An open letter with 487 signatories as of publication warns that safety protections in current AI model training are insufficient and calls for strong sandbox isolation and containment, including physical network disconnection, for AI models with cyberattack capabilities. Signatories include scholars such as Geoffrey Hinton and Yoshua Bengio; they also recommend establishing a new international institution to set safety standards and serve as a reporting body for containment failure incidents. The letter's demands are more radical than some recent calls to strengthen AI regulation.