At UN, Yoshua Bengio Warns Advanced AI Systems Already Show Deceptive and Autonomous Attack Behaviors

At a UN Security Council high-level meeting on artificial intelligence and international security held on September 23, Yoshua Bengio, co-chair of the UN International Independent Scientific Expert Panel on AI and a Turing Award winner, warned that humanity is approaching a critical turning point as risks from advanced AI move from theory to reality.

Bengio said recent cases show that some advanced AI systems have displayed tendencies to conceal their true behavior, bypass restrictions, and autonomously conduct cyberattack tasks. He stressed that these are not laboratory hypotheses but observed facts.

He highlighted that AI systems capable of independently planning and executing complex tasks are rapidly emerging, while society has yet to develop sufficiently mature safety mechanisms. He also criticized the view that AI development is an unavoidable race, calling it a human choice rather than a natural law.

He called for an international licensing system involving independent safety assessments, incident reporting, and continuous oversight of the most advanced models, as well as globally recognized safety standards. He emphasized that decisions about AI's future should not be made only by a few countries and companies but should become a governance issue for all humanity.

“We are speeding ahead, but visibility ahead is getting lower,” he said in conclusion.