I've been thinking about the Hugging Face hack by the botched Open AI test and I'm sure others have reached this conclusion, but I haven't seen it explicitly stated. If swarms of AI agents can successfully and feloniously penetrate systems of a private company in an exercise meant to be contained...