When English mathematician Matthew Watkins went to San Francisco to study an early version of ChatGPT, he expected dry technical work. Instead, he uncovered a series of 'rare tokens' that made the system go haywire. Some produced gibberish.Others turned it hostile. One trigger word - petertodd - sent the machine into what looked like full-blown paranoia. Was it a simple bug? A trail of breadcrumbs left by whistleblowers? Or a glimpse into how AI really thinks?The search for answers pulls Watkins into the secretive world of AI safety research.In this rapidly expanding domain, the line between science and belief blurs: billionaires play god while rogue chatbots win fortunes and groupies. What starts as a computer mystery becomes something much bigger: both a window into the eerie inner worlds of LLMs and a cryptic clue to what might be history's greatest power grab.The system has since been patched. Even so, before the window closed, we saw inside for a brief moment - and what stared back at us was both stranger and more human than we might like to admit.