4096-bit encryption sounds great, but there are always ways to shave orders of magnitude off of the actual sample space, such that encryption strength really tends to grow at about the same rate as processing power.
Once you get encryption keys of that size, you've got storage and transmission issues, which increase the probability of other attack vectors working. Plus, your PRNG has to be REALLY random -- and there isn't really all that much true entropy to go around when you get right down to it.
The issue here is that as your random seed gets larger, the probability that it isn't truly random also increases, and analysis of data encrypted with this seed becomes easier through replay analysis. It won't remove the actual entropy, but forces acting on the values generated will create patterns that will still limit the amount of true entropy stored in the resulting value.
Kind of like if you flip a coin once, anyone guessing really doesn't know if it'll come up heads or tails. But if you flip a coin ten times, the method you use for flipping the coin and the environmental factors will start to have an impact on which side comes up more often, and also on the pattern of what influences a heads or tails result. If you flip the coin 2^4096 times, you'd probably be able to pretty accurately predict the result before the coin had even fully launched into the air.
Yeah; the coin experiment is often used as an example of how entropy is entropy and the probability doesn't change from toss to toss. But if you take all other factors into consideration, you limit the effect of entropy such that your guess on a given toss can actually improve over time. Try it, and you'll see I'm correct.