(This is aside from the fact that we're talking about an extremely elaborate way to construct a CPA attack against a protocol that does not automatically give attackers chosen plaintext).
A slight digression: Asking how you'd actually figure that out, in the context of arguing for the security of the system, is the wrong question and a flawed approach to designing cryptographic standards. The same question could be asked about the last packet IV issue Rogaway raised. No one knew at the time how to exploit it then, but it was still a bad idea, as time has proved. I can't see the attack is not the right design philosophy for cryptographic system. To anyone reading this, please please don't use it. It usually leads to broken shit and, if some of the people proposing the standards are malicious, certainly will if they are cleverer than you ( and the NSA probably is).
Now, if you are just curious how to break it: All you have to do is get a few IVs. Suppose the system does encrypt then mac. Let's assume you can get the system to decrypt its own messages and give you the result (this is a weaker version of a chosen cipher text attack and authenticated AES CBC is still safe even with this issue provided IV's are random). Simply replay a given message a few times. The IV changes each time, and hence you get the result of m xor IV_1, m xor IV_2, m xor IV_3, etc. Minimally, this gives you the XOR's of the various IV's with each other which may be enough to calculate the LFSR state. If you happen to know m as well, then you have the IVs. Game over.
Lastly, assuming you don't get a chosen plaintext attack for a protocol with as broad uses as IPSEC seems like a bad idea. Minimally, you clearly get one if it's used to secure http with javascript in it via the same mechanisms as BEAST. You also get one in the case of ICMP echo. File sharing protocols would seem to also readily provide them.
[0]http://crypto.stackexchange.com/questions/5293/cryptanalysis...
I don't think LFSR CBC IVs are a good idea (though again note that they work in CTR), but they seem like an awfully dumb way to backdoor a standard, and it's straightforward to see how a well-intentioned suggestion could propose their introduction.
1) AES CTR needs a nonce (number used once) not an IV. So a maximal LFSR works up until it repeats. People call it an IV, but strictly speaking IV's need to also be random.
2) I have no idea if you need the replays in a very specific order or not. If you know the m being decrypted, then you always get the IVs. I suspect if given random parts of the output sequence of an LSFRs and the distances between them, it's still easy to recover the seed.
Then it's just a matter of mounting an (adaptive) chosen plaintext attack against the system knowing what the next IV will be. This is something you can't do against AES CBC mode if the IV is actually random.
It fails in the same way if you use the last packet's cipher-text as the IV and there are pauses between packets.