So what the researchers did, apparently, was to gather all the RSA public keys they could find (6 million or so) and then calculate the gcds of all pairs of keys. (Presumably they were slightly more clever than this, but it doesn't really matter--doing this is still vastly easier than factoring even a single public key). Whenever they found a gcd that wasn't equal to 1, they'd cracked (at least) 2 keys. And it turns out that roughly 0.2% of the keys could be cracked in this way (for a total of 12720 1024-bit moduli).
As far as I can tell, the problem isn't with RSA itself, but with whatever algorithms are used to pick two random primes to multiply and produce the key (or moduli, whatever)--they don't have high enough entropy and output duplicates ~0.2% of the time.
But as far as I can tell, the result is that, of public keys collected "in the wild" show more duplication than should be expected, so the random number generation schemes used are not "random" enough. Is that it, or is there more to the story?
(Edit) No, there's more. I don't quite understand (and don't think they really explain), but somehow they use the fact that there's lots of keys to find ones with common factors, which allows you to factor both of them. This sounds basically like the Euclidean algorithm but I couldn't tell what method they're using to find them.