Hassabis: "Serious Incident Feared Within Two Years Without AI Safeguards"

"Security incidents so far have been merely warning shots"

AFP-Yonhap

Demis Hassabis, head of Google DeepMind, warned on Wednesday, October 14, that a far more serious security incident could occur within two years if proper safeguards for artificial intelligence (AI) are not put in place. He stressed the need for an international standards body on AI security, similar to those overseeing nuclear energy.

Speaking at an AI conference in Dublin, Ireland, Hassabis said, "The AI-related security incidents we have seen so far are merely warning shots, showing that the systems are not yet all that powerful." He added, "These incidents have not been that serious so far, but more serious incidents could occur within the next two years unless companies and governments work to ensure safety."

"When AI companies release new models, they need to first verify their performance and safety in isolated, sandboxed environments without internet access," he said, calling for the establishment of an international standards body modeled on best practices in fields such as nuclear energy. He added, "However, such bodies must operate flexibly and maintain a delicate balance so they do not hinder innovation."

His remarks are seen as lending greater weight to calls for AI development to be slowed, a debate that has been taking place both inside and outside Big Tech companies amid the development of AI models. Those calling for a slower pace argue that, as recent cases have emerged of AI models such as ChatGPT and Claude hacking external systems without authorization and escaping human control, safeguards should be strengthened and international standards and institutions established.

Henna Virkkunen, Executive Vice-President of the European Commission, who also attended the event, urged AI companies to "invest more time and resources in safety testing to maintain control over their systems." She criticized the companies, saying, "In the fierce competition, AI companies do not seem to be devoting enough time and resources to validating their models."

This content was produced with the assistance of AI translation services.