Using Ordered Markov Chains and User Information to Speed Up Password Cracking

But an RNN isn't necessarily going to help as much as you think. An RNN has two problems compared to a Markov chain:

1. Markov chains memorize strings very very easily, accurately, and scalably; it's easy to memorize phrases, words, suffixes, and prefixes from the existing corpuses of billions of passwords. That's all a Markov chain does, memorize & count. On the other hand, an RNN will struggle to do so because there's no 'place' for it to put all of that, everything has to be encoded into the fixed set of neural net weights, otherwise, it just doesn't know about it; and the more you ask it to learn, the more the competing demands fight each other. RNNs augmented with external memories might help fix this but are still cutting edge research.

2. Markov chains are also very fast, far faster than an RNN. Multiple orders of magnitude difference are possible, unless you use a RNN so small as to be irrelevant (since then it can't memorize anything). For cracking hashes, a small gain in plausibility of guesses is not worth being able to make hundreds or thousands times fewer guesses (unless perhaps the hash are something proper like bcrypt/scrypt where it takes seconds to check, in which case the guessing phase takes up a much smaller fraction of runtime and better guesses may be worthwhile).