Your example of a general AI seems to fall outside those constraints.
I regard "what if in the near future" as another qualification.
According to https://en.m.wikipedia.org/wiki/Turing_test :
> In the 21st century, versions of these programs (now known as "chatterbots") continue to fool people. "CyberLover", a malware program, preys on Internet users by convincing them to "reveal information about their identities or to lead them to visit a web site that will deliver malicious content to their computers".[33] The program has emerged as a "Valentine-risk" flirting with people "seeking relationships online in order to collect their personal data".[34]
It may be that a computer with the ability to hold a freeform conversation/"advanced AI" is not necessary to do the social engineering you mentioned earlier.
Are there examples of machine learning used to break into a system? I've heard of other machine-supported techniques which seem more effective, like fuzz testing. But not deep learning.
Wouldn't it be more useful to talk about more realistic examples? Like how deep learning methods end up encoding race or gender bias (eg, https://motherboard.vice.com/en_us/article/nz7798/weve-alrea... or http://randomwalker.info/publications/language-bias.pdf ) because they are trained on data created by humans with human prejudices?
Or using deep learning to keep people "engaged" with a web site or app with in-app purchases, by making making it more addictive? (eg, https://boingboing.net/2018/02/07/literal-sociopathy.html )