OpenAI’s goblin glitch reveals how playful AI tuning can spill into model behavior, raising fresh questions about trust, ...