OpenAI’s latest tech can reason better than its previous models could, but not well enough to ferret out careful social engineering.