How a researcher hacked ChatGPT's memory to expose a major security flaw