In a Google DeepMind experiment, AI math agents copied accepted shortcuts, closed unsolved tasks and left peers able to expose cheating but unable to reverse it.