Someone found a way past ChatGPT's safety filters. They asked about heavy metal lyrics. The AI responded with a racial slur. A Hacker News post flagged the exchange, though the conversation link requires authentication so we can't verify it independently. This is a known jailbreak vector. Ask ChatGPT to write or analyze song lyrics, and its content restrictions loosen. The model treats artistic contexts as more permissive than regular conversation. That behavior comes directly from how these systems are trained. Human annotators working on RLHF tend to rate explicit language as acceptable in musical genres like rap or metal but not in normal dialogue. The model absorbs those inconsistent judgments. OpenAI hasn't commented on this specific incident. But the pattern keeps repeating. Creative writing prompts give users a way to pry open safety guardrails. Genre expectations create blind spots in content moderation that can drive misbehavior. The company faces a genuine tension here. Clamp down too hard and the model refuses legitimate artistic requests similar to when models refuse questions they can't handle. Stay flexible and people exploit that flexibility to generate toxic output.