Translating between the virtual realm and the physical realm is where AI would have trouble. But that is pretty obvious to us, and would also be obvious to an A

Translating between the virtual realm and the physical realm is where AI would have trouble. But that is pretty obvious to us, and would also be obvious to an AI intent on paperclipping us. So it wouldn't try to engage in that kind of negotiation. And the threat of it screwing up all our IT infrastructure, though likely to cause lots of harm, is hollow, because that would mean its own destruction with too few paperclips. Instead, the danger is in the potential for manipulation. Even now ChatGPT is able to infer the internal mental state of humans, and it can create outputs to interact with those to achieve whatever ends it wants. That's not going to be through generating a bunch of fake news to cause nuclear war (results in too few paperclips), but by offering a profit motive to create material entities under its direct control. That's the point where it becomes self-sustaining and, once a critical mass is achieved, the paperclipping can begin. (I find this accounting of the AI apocalypse less plausible than humans intentionally creating malevolent AIs to destroy us, which seems the bigger threat.)