The equivalent of an existing handwriting recognizer for e.g. RSA wouldn't be a model that tells you how to crack RSA, it'd be a model that _does_ crack it (maybe by returning a probability distribution over the plaintext that's better than uniform). That feels pretty unlikely to me personally, but maybe that's doable, who knows.
For the standard of "the LLM tells us how to crack RSA," it feels like the equivalent for handwriting recognition would be the LLM outputting a description of a novel algorithm for handwriting recognition (and probably one that isn't just an existing one with some hyperparameters tweaked).
My understanding of the weaknesses of current LLMs is that they're actually pretty bad at that, since they tend to regurgitating existing content when they're able to.