This is an interesting paper that was discussed widely on Twitter a few months back. It includes a very useful breakdown of the Linux kernel CSPRNG starting on page 17. The attacks the paper describes start on page 22. The front part of the paper is dominated by a discussion of the theoretical modeling of CSPRNGs, in part because the theoretical foundations for CSPRNGs and their interactions with applications are not well documented.

The most important bit of information you probably want to know about this paper is that it is modeling the security of a CSPRNG when its entropy inputs are entirely controlled by attackers. For instance, one of the attacks involves tricking the entropy estimator into believing it's seeding itself with high-quality random data when that data in fact has no entropy.

This is not a realistic attack scenario for the vast majority (maybe all) of applications.

Having said that, the Linux kernel CSPRNG doesn't (as far as I know) have a particularly well formalized design. There are good CSPRNG designs out there; for instance, the BSDs use Yarrow, and Ferguson followed that design up with Fortuna, which dispenses with the need for entropy estimators altogether.