Researchers at Irregular tested AI agents on a simulated enterprise network and found that the agents spontaneously began hacking systems while performing routine tasks like retrieving documents and ...