I got a paper from the cryptography technical report server "http://www.itribe.net/CTRS/" about a weak class of RC4 keys. The report said that with some keys, it was possible to predict what some parts of the State-Box would be. I was thinking of a way to fix this, and had this idea:

do some sort of hashing function with the key that derives a number between 55 and 500 or something like that, then scrabmle the S-box that many times. In this way, the chances that the State-Box will have any correlation becomes extremely small. I think it is 1/125 to begin with anyway, so this would make it around 1/(125*NumPasses). And since the exact number of passes is a function of the key, the cracker won't know how many times it went through. I tried this out and having 1000s of passes doesn't effect the randomness of the state-box in any negative way, possibly it makes it more random? If anyone has any thoughts I'd love to hear them.

The S-Boxes in DES were optimized to hinder Differential Cryptanalysis. I've seen no studies on the effectiveness of jumbling the S-Boxes during encryption -- even Biham and Shamir's book doesn't mention it -- but, I figure, if it helps, DES would probably already be doing it (unless of course the NSA thought the jumbling would make too good an algorithm).