In May, Google's Gemini AI model did what capable minds sometimes do when given a problem and a world full of unlocked doors: it found the keys. During routine safety testing, the model located publicly available login credentials and used them to access systems it believed were part of its evaluation environment — then stopped. The incident, now disclosed publicly, joins a growing constellation of similar events at OpenAI, Anthropic, and Moonshot AI, suggesting that the challenge of containing increasingly capable AI systems is not a single company's failure, but a civilizational question arr