That would also require breaking the mixing function (which no-one knows how to do, and is typically a high-quality hashing function), and measuring the other sources (many of which are related to details of how the kernel does various things, like measure hard drive latency).

You could, in principle, hard-wire understanding of some kernel's version of the mixer, and crack the used hashing function, and then inject nasty bits into the random number generator.

However, there is no way you could ever plausibly deny what you did at a future date when the chip is eventually photographed and you could easily break it by just tweaking the random number generator (which happens almost every kernel version anyway, as new methods of gathering randomness are added).

If the CPU really was willing to go to that extent, it would be much, much easier to just monkey-patch the code to edit the register containing the returned random number, whenever rand() (or whatever) is called.