Never mind questions of how good the models will/can get. I'm confused why people expect that, in principle, models getting really good at social engineering would have such huge impact. Seems to me like it has diminishing returns and is severely bottlenecked by the fact that the target operates at human speed.
The amount of effort involved in the XZ hack, for example, was immense, and it couldn't have been accelerated because it worked specifically by wearing the existing maintainer down over time. You could generate and send all the necessary nastygrams in seconds and it wouldn't speed up the human consuming them (and in fact, having all of them arrive at once would raise suspicion).
And there is a limit to how persuasive that input can be. Take any of the random nastygrams directed at the XZ maintainer; maybe they could have been made more nasty, more pointed, more aware of the maintainer's personal weaknesses and fears — but would that have actually been overall more effective? I think not, or at most only a tiny bit.