Long-time readers may have noticed that spam on the wiki has been a very persistent problem for the past 2 years or so; I've been dealing with it so far by hand, but I recently reached a breaking point and asked Trike to resolve it or find a new wiki administrator. (Speaking of which, is anyone interested?)
So Trike has enabled a MediaWiki extension called the edit filter: a small functional programming language which lets you define predicates applied to edits which trigger one of a set of actions, like banning a user, deleting an edit/page, or stopping an edit from going through. I have so far defined one rule: page creation is forbidden for users younger than 24 hours. This so far seems to have worked well; spam pages have fallen from 5-10/day to ~5 over the past 2 weeks. This is much more manageable, and I am hopeful that this new anti-spam measure will be effective longer than the previous additions did (but if it doesn't, I'll look into adding more rules dealing with images and external links, and perhaps also ban users whose names end in a numeric digit as almost all the spam accounts do).
If you've run into this edit filter before by making a page and seeing the submission rejected with an error message, fret not: merely wait 24 hours. (If your account is more than a day old and you're still getting errors, please contact me or Trike.)
I'm not worried. As you can see from the sidebar, spammers have been prolifically creating accounts for countless months and almost all accounts wind up never being used. My inference is that most of them are being stymied by other anti-spam features. None of the spam seems to be done by hand, and certainly they aren't looking on an obscure post on a different domain for a value that they would stumble upon accidentally ('hm, my spam account didn't work yesterday, let's take a look to see if the error went away - oh, it did, there must be a 24-hr timeout').
I totally agree there's a very low probability of leak there. There's still a (meta) reason to do stuff like this though.
If you have a high value target (like building an AI), you need insane paranoid security. Means you need to train the mindset of patching everything you possibly can not because you think it's unsafe but, you know, just because. E.g. you could tell the difference between good and bad sysadmin just by looking at his own PC. A good sysadmin will have every single drive truecrypted, no matter the inconvenience. This is an opportunity to train the mind into the job.