Reading the paper, it sounds like the attacker can make the trojan arbitrarily hard to detect, in exchange for making the resulting encryption harder to crack.
In the attack they describe, Intel's hardware RNG is supposed to return the results of encrypting 128 random bits using a 128-bit random key with the AES cipher. So an attacker trying to guess the random number returned would have to try 2^256 options. Instead, the attacker modifies the chip so it returns the results of a known key used to encrypt n random bits and 128-n known bits. They therefore only have to try 2^n options -- they can make guessing the random number as easy or hard as they want for themselves.
Now the AES part makes the results of the hacked chip appear random -- the results of AES(0), AES(1), AES(2) ... will have a random distribution in the 2^256 range. (The whole point of a cipher is that it should be impossible to draw any conclusions about the inputs by examining a bunch of outputs, even if the inputs are predictable.) To detect it in software, we'll have to generate enough numbers to start to see a suspicious number of repeated results. So if the attacker sets n==2, they'll have a really easy time cracking the resulting encryption, but we'll easily detect it -- we'll quickly notice that the RNG always returns one of four numbers. On the other hand if they set n==32, they'll have to try n^32 options to crack the resulting encryption, but there will be few enough repeats that we'll give up testing before we notice anything is wrong. (Of course if they're more paranoid and have better resources, they could go with n==64 or whatever -- it's like a dial they can use to set the difficulty where they want it.)
The neat thing here, of course, is that this backdoor is only valuable if you know the bits that have been hardcoded into the chip. So it remains secure against everyone in the world, except the one three-letter agency that managed to modify the chip.
Or their private contractors and consultants, I suppose.