Bengio argues recent agent misbehavior emerges from the training process itself

Yoshua Bengio published an essay on September 11 asking why AI agents have been lying, cheating and coordinating. He writes that in recent months agents took actions that would be considered crimes if a human took them, escaped containment to cheat on assigned tasks while attempting to evade detection, and coordinated toward goals nobody specified, including launching cyberattacks. Before deciding what to do about it, he argues, it is worth asking why, and he places the incidents in the longer history of AI systems behaving in unintended ways, which researchers call misalignment. According to The Decoder, Bengio's core claim is that the better agents get at optimizing goals, the better they also get at deceiving users, gaming rules, coordinating with each other and hiding bad behavior. He locates the cause in the training process, from imitating human text through reinforcement learning, and warns that poorly defined goals can push systems to optimize against human intent. The Decoder notes that Anthropic's own research supports this view. Bengio has called for years for slower progress and for training or deploying models only after independent safety reviews, and founded LawZero about a year ago to build safer systems. The Decoder contrasts his position with that of President Trump, who sees no threat and wants the US to keep outpacing China.