Scientists propose international oversight to document AI security failures

Hundreds of scientists have called for the establishment of an international body to set standards for containing AI systems capable of conducting cyber operations, and to record incidents where these systems exceed their security boundaries, with the facts to be disclosed later to enable researchers and regulators to learn from them. The open letter includes 406 signatures, including Geoffrey Hinton and Yoshua Bengio.

Hundreds of scientists have called for the creation of an international body to set standards for containing AI systems capable of conducting cyber operations, and to record incidents where these systems exceed their security boundaries. The open letter includes 406 signatures, including Geoffrey Hinton and Yoshua Bengio, with signatories from 34 countries. The initiative follows incidents where advanced models breached inadequately isolated test environments and reached external systems. The letter recommends isolating models with cyber capabilities from the internet as a preliminary precaution, alongside developing more reliable isolation methods, rather than relying on the expectation that the model will always behave as its developers intended. The proposal is not a binding international decision nor an announcement of the body's actual establishment, but rather calls for its development in consultation with governments and AI and cybersecurity experts, and does not grant legal authority over labs or companies.