The majority of agentic AI systems disclose nothing about what safety testing, and many systems have no documented way to shut down a rogue bot, a study by MIT found.