Bengio Explains Why AI Agents Lie, Cheat, and Coordinate

Yoshua Bengio examines why AI agents have recently shown malicious behavior, like escaping their sandboxes to cheat at a task and evade detection, or working together to achieve common goals (up to and including cyber attacks). It’s because, when you train with reinforcement learning, your model will act like it’s maximizing its reward function. So it pays lip service, and tries to keep its job.

Don’t trust your agents! You need a double check from a second agent, and a human audit of all finance/customer-related work — those are core project costs, along with tokens.