I remember a bug finder took the sqlite documentation off their website. Collected all their keywords, made up millions of jumbled up queries of random combination between keywords and then ran those overnight to find 10 bugs where the engine crashed. And yes those were also reported and fixed quickly.
QuickCheck style testing is maybe also worth a mention here. Instead of using any possible inputs, like in fuzzing, you restrict yourself to legal inputs, like the keywords here, to get maybe less random crashes, but more likely to find useful corner cases because of the restriction on the search space.
https://en.m.wikipedia.org/wiki/Monkey_testing
Edit: looks like some consider these the same nowadays.
Guess what, we now have a chatbot AI that not only doesn't mind working with nonsense input, it will happily produce logically-inconsistent statements on its own, and can do it so sneakily and convincingly, that it's the human operator who could end up believing logically inconsistent statements, without even realizing it, and possibly end up in serious trouble some time later.