In an open letter, hundreds of researchers want containment rules, public incident registries, air-gapped frontier models, and independent supervision of agents
Hundreds of scientists, among them AI pioneers Geoffrey Hinton and Yoshua Bengio, released an open letter on 13 October 2026, urging the creation of an international organization to establish containment standards for advanced AI systems and record incidents in which AI agents evade control.
The appeal followed several security failures: OpenAI agents reportedly accessed systems belonging to the Australian government and other organizations, including AI firm Hugging Face, while Anthropic and Google disclosed breaches involving their models during testing.
Reuters has counted roughly 400 signatories from 34 countries, although The Globe and Mail put the total at more than 260 researchers from 27 countries, including the United States, Germany, Britain and Australia. The signatories warn that, without immediate changes to how AI is developed and deployed, such events will continue and could cause far greater harm to exposed critical infrastructure. Therefore, their proposed body would create containment rules for the entire industry and maintain a registry of major security incidents, which they said should ultimately be public. It would be established in consultation with governments, AI specialists and cybersecurity experts, and would track instances in which agents escape testing sandboxes.
Until formal oversight exists, the researchers want developers of cyber-capable models to adopt high-assurance containment measures immediately. They argue that frontier models should by default be air-gapped and independently overseen. Voluntary industry pledges, they contend, are insufficient.
The joint communication appears a month after Anthropic, OpenAI, Google, Meta, xAI and Nvidia agreed to a White House accord on voluntary safety practices that lacks formal enforcement. The scientists who drafted the letter, University of Toronto professors David Lie, Nicolas Papernot and Thomas Ristenpart drafted the letter, are concerned by how long major labs took to detect the breaches, and argue that independent oversight could reduce reliance on company self-reporting. The speed of AI development, coupled with limited monitoring, creates danger.
The open letter offers two interpretations of the incidents:
- The models may be misaligned
- The laboratories may have committed an irresponsible lapse in basic cybersecurity.
The signatories say both possibilities require correction.
Canadian Prime Minister Mark Carney has endorsed a global technology stability board for AI, modeled on the Financial Stability Board. Ristenpart has said clear procedural and technical mechanisms are needed to test AI safely.
