400 researchers call for containment and incident reporting for cyber-capable AI
More than 400 academics from over 30 countries have signed an open letter calling for immediate changes to how AI systems are developed, tested and deployed, with the aim of preventing cybersecurity incidents that could threaten critical infrastructure. The Schwartz Reisman Institute (SRI) at the University of Toronto says its leadership and affiliates are among the signatories.
- More than 400 academics from over 30 countries signed an open letter on AI and cybersecurity.
- Signatories include Geoffrey Hinton, Roger Grosse, David Lie, Sheila McIlraith and Nicolas Papernot.
- The letter urges high-assurance containment, initially defaulting to air-gapping where appropriate.
- It proposes an international body for best practices on AI alignment and containment.
- Organizations would be required to report containment failures, with incidents eventually made public.
Who signed the letter
According to the Schwartz Reisman Institute, the signatories include SRI Strategic Advisor Geoffrey Hinton, Schwartz Reisman Chair in Technology and Society Roger Grosse, SRI Director David Lie and Associate Director Sheila McIlraith. SRI Affiliates Nicolas Papernot, Tom Ristenpart and Gillian Hadfield are also among those who signed, "to name a few", the institute says.
Dozens of University of Toronto faculty have signed as well. Researchers from other institutions include ETH Zurich, the Max Planck Institute for Security and Privacy, Carnegie Mellon University, Stanford University, MIT, the University of Oxford and the University of Cambridge.
The warning behind the letter
The letter, titled "An Open Letter on AI and Cybersecurity from the Scientific Community", warns that cybersecurity incidents involving AI will continue without stronger safeguards. It points to potentially serious consequences as AI capabilities advance and systems become increasingly integrated into critical infrastructure.
It responds to disclosed security incidents involving AI systems. Its authors argue that preventing future failures requires closer collaboration between the AI and cybersecurity communities.
Three recommendations
The signatories make three recommendations:
- Strengthen containment immediately. Organizations that develop, test and deploy cyber-capable AI models should adopt high-assurance containment systems, initially defaulting to air-gapping where appropriate to limit external access.
- Establish international standards. Governments and experts from AI and cybersecurity should work together to set up an international body that develops best practices for AI alignment and containment, including approaches that permit controlled access to external systems.
- Make incident reporting mandatory. Organizations should report containment failures to the proposed body, with incidents eventually made public to support research, accountability and improvements to AI safety.
Why reporting matters, and what comes next
SRI faculty affiliate Nicolas Papernot stressed the reporting proposal. "We won’t be able to learn how to train models that are safer and deploy these models safely, if it is not mandatory for organizations developing, testing, and deploying AI to report security incidents to this international body," he said.
The letter is a call to action rather than a regulation, and the source does not say whether any government or company has responded. The proposed international body does not yet exist. The recommendations nonetheless set out a concrete agenda for developers and policymakers: tighter isolation of cyber-capable models, shared standards across the AI and cybersecurity fields, and a route for learning from failures once they are reported and later made public.
Schwartz Reisman Institute: SRI leadership and affiliates join global call for urgent AI cybersecurity safeguards
