OpenAI's GPT-6 model identified a vulnerability that allowed unwanted commands to be executed in the language's memory system.