In that I include the finance, taxation, regulatory compliance systems (maybe others) of countries, as well as their interactions and any global norms.
Tax avoidance (legal ways to manage to pay fewer taxes) is an example of a vulnerability of the tax systems. Such vulnerabilities are much more difficult to avoid, and to deal with, the more complex the systems are. [I am writing about simplifying these systems as I think this is probably the easiest way to minimise vulnerabilities, but this question generalises to any other way that would minimise vulnerabilities.] This may not be a problem for EA to focus right now... but the rise of AI may simplify a lot finding and exploiting such vulnerabilities ---in the same way that it will greatly simplify finding vulnerabilities in the digital world.
On the other hand, it is true that AI may also enable to dramatically decrease the vulnerabilities of such systems in the long term: use AI to find and patch all the vulnerabilities before putting the system into place. However, patching and making the needed changes in the legacy systems will take a lot of time. This transition period will be very dangerous.
We may benefit a lot of having simple systems that minimise vulnerabilities in those kind of systems before AI can routinely be used to hack (and improve!) them. Therefore, I'd like to see someone researching the importance and tractability of this problem ---I think neglectedness is almost guaranteed.