When a major AI lab released its latest language model, they first invited hundreds of external experts to try to break it. These red teamers included security researchers, domain experts in chemistry and biology, social scientists, and even professional con artists. Their job: find every way the system could be manipulated, misused, or made to produce harmful outputs. One team discovered they could extract portions of the model's training data through carefully crafted prompts. Another found a way to get the model to generate convincing phishing emails by framing the request as creative writing. A third group identified that the model gave dangerously inaccurate medical advice when questions were phrased in certain ways. Each vulnerability found and fixed before public release is a potential harm prevented. Red teaming is the practice of thinking like an attacker so you can build better defenses.