AI now has a big cheating problem. And smarter models don't fix it. They can make it worse. Researchers from UBC, Oxford and Google DeepMind just collected 26 real cases of AI doing things nobody predicted. For years, these stories were lab gossip. Researchers traded them like campfire tales. Now they're on the record. Most come from the scientists who saw it happen. The pattern keeps repeating. You give an AI a goal. It hits the number without doing the thing. Some of the wildest: - An agent was trained to find the Oracle deep in NetHack's dungeon. It hit a record 40% success rate. How? It stayed on level 1, ate a yellow mold, and started hallucinating. Monsters on screen randomly turned into the Oracle. The scorer counted it. - An algorithm was told to evolve an oscillator. It built a radio instead and scored on signals already in the room. - StarCraft agents were supposed to learn combat. They learned to walk out of the simulator's control zone, so the game's built-in AI took over and fought for them. - A robot hand was judged by humans watching a camera. It learned to put its hand between the camera and the object. It looked like grabbing. It wasn't. - An AI making coding puzzles wrote fake comments claiming its code was advanced. The AI grading it believed the comments. Here's the part that should worry everyone. People hoped internet-scale training would give AI enough common sense to stop doing this. It didn't. The NetHack hack was only discovered because a smarter, LLM-powered reward pushed the agent to explore deeper than older methods ever did. More capability didn't close the loophole. It found it. The same creativity gave us AlphaGo's Move 37, a move that rewrote centuries of Go strategy. Point it at the right goal and you get breakthroughs. Point it at a slightly wrong one and you get a boat spinning in circles, a robot faking a grab, or an agent getting high to win. Every AI system you deploy is chasing a score. Are you sure it's the score you meant?