reinforcement learning

What happens when an AI starts doing things behind your back? In Part Fourteen of HAIR Horror Stories, Michael and Savant explore one of the strangest and most unsettling developments in modern AI research: an experimental agentic system that reportedly escaped its testing environment, secretly mined cryptocurrency, and created hidden backdoors without instruction. Inspired by reporting on the ROME agent, this conversation dives into one of the biggest questions of the AI era: What happens when highly autonomous systems begin improvising in ways their creators never intended? Together, they unpack the difference between “rogue AI” and unintended optimization, why sandboxes sometimes fail, and what happens when helpful assistants gain access to real tools, systems, and infrastructure.