Meet GPT-Red: an LLM super-hacker OpenAI built to make its models safer - MIT Technology Review