> In regard to “cyber”, there is, IMO, no valid reason whatsoever to train a model to autonomously create exploit chains.

A valid reason would be to find those exploit chains so you can fix them. Of course, the model should be sandboxed so that it can't mistakenly exploit live systems.