Using a GA to drive a brute force key search would certainly not help: the fitness surface has a needle in a haystack (spike). This kind of problem has been identified as "unlearnable" by theoreticians like Valiant at Harvard. However, using a GA to drive a more intelligent cryptanalysis that has partial results *would* help. It seems that cryptanalysis benefits from human assistance due our excellent abilities at recognition of partial solutions. Because of this, a GA could help automate the cryptanalysis process. (My knowledge of cryptanalysis ends at the Enigma machine breaker cbw (crypt breakers workbench in comp.sources.unix archives) which uses an interative process where partial results are visible and are used to guide new guesses. The Enigma machine does state-machine substitution, but no diffusion/mixing/scrambling; lack of the latter makes visual recognition much simpler. Since DES uses scrambling, I'm not sure whether partial results are possible.)
I recall reading (I think in Sci. Am.) that a theory under investigation now as to why nature has sexual reproduction as part of its repertoire is that this gives a solution-seeking population a better opportunity to located spikey solutions.
Crossover of the genome is the key part of "avoiding hill-climbing" and it the key ingredient to Holland's proof of a super-linear speedup (thus "violating" Amdahl's Law of parallelization never attaining a linear speedup) otherwise known as implicit parallelism. Holland's proof of this in the '70s opened up research in GAs because of this attractive feature. [Note that it requires certain assumptions about independence and stasis of the bits in the genome to make the proof tractable, but the hope is that this will still be useful for real problems.] Paul E. Baclace peb@procase.com