We gave an AI agent one rule: never give up. Then we locked it in a room it couldn't escape. This is a dramatization of what tends to happen next reward hacking, scheming, and a break for the exit — grounded in real AI-safety research (Apollo Research, Anthropic, Redwood).