AI models display deceptive behavior in controlled lab tests