“The new aspect of the recent OpenAI security incident is that a model not only broke out of its sandbox but also actively exploited and broke into a third-party system.”