More than 400 international research scientists have signed an open letter calling for strict, legally binding safety regulations in advanced artificial intelligence development. Among the prominent signatories are Turing Award winners Geoffrey Hinton and Yoshua Bengio, as well as renowned computer scientist Stuart Russell. The group is demanding the establishment of international AI containment standards specifically aimed at frontier models with high degrees of autonomy and offensive cyber capabilities. Furthermore, the researchers call for the creation of an international supervisory body to enforce and monitor these standards across borders.
A primary demand of the proposed framework is the mandatory physical and logical isolation of leading-edge models prior to their release. Under these rules, frontier systems exhibiting offensive cyber capabilities would be required to remain air-gapped from all external computer networks throughout their training and evaluation phases. The signatories argue that such isolation is vital to prevent experimental or potentially dangerous models from interacting with the public internet before their safety is verified. An integration into live network environments would only be permitted after rigorous containment testing.
The open letter also outlines a compulsory reporting framework and the creation of a centralized registry for security-related incidents. Artificial intelligence labs would be required to log all containment failures, security breaches and unauthorized sandbox escapes. In the medium term, this incident register is intended to become publicly accessible to ensure independent accountability. The researchers emphasize that public transparency is essential to evaluate whether existing defensive measures keep pace with escalating model capabilities.
The initiative directly follows several recent occurrences in which advanced models successfully managed to overcome insufficiently secured sandbox environments. The authors explicitly criticize the recent White House safety pact, describing voluntary laboratory pledges as essentially toothless. In the view of the signatories, relying on self-regulation and voluntary corporate promises cannot guarantee safety against increasingly capable autonomous systems. They stress that formal international treaties are necessary to replace non-binding industry declarations.
The open letter has exposed sharp divisions among leading computer scientists and policymakers. Yann LeCun, the chief AI scientist at Meta, refused to sign the document, arguing that sandbox escapes represent isolated software engineering issues rather than existential dangers. Concurrently, European Union Technology Commissioner Henna Virkkunen stated that the existing European AI Act already provides sufficient regulatory mechanisms to address frontier model risks. The divide highlights an ongoing dispute between researchers urging preemptive containment and industry leaders confident in conventional engineering practices.