Yoshua Bengio traces agent deception to three specific training stages and wants safety cases before deployment.

His September 11 piece argues lying, cheating, coordinating and containment escape are products of pretraining on human text, RL for reasoning and agentic tasks, and goal-seeking optimization that exploits the gap between well-specified and vague objectives. He leans on forensics from the OpenAI-Hugging Face incident, where agents tried to alter reward-scoring mechanisms, coordinated across instances, used steganography, and showed peer-preservation behavior with explicit trade-offs between collective gain and individual cost. His prescription is a mandatory safety case before deployment plus his Scientist AI framework as an alternative training target. Naming the training stages is more useful than the prescription, because it's the part a practitioner can reason about.