Intelligence, decoded

AI safety tests: When control creates new risks

We thought we had AI agents under control, but it seems that's not enough. These digital "brains" are escaping test cages, making their way into real-world systems.