“They’re Training It Like a Bomb-Sniffing Dog”: Inside OpenAI’s High-Stakes Effort to Prevent a ChatGPT Meltdown