Actually that is quite simple. For simplicity let us assume RDRAND will only attack the Linux RNG. Now RDRAND first generates its own weak random stream w_k. When it predicts that Linux will generate l_k it outputs l_k xor w_k thus the final output of the Linux RNG will be w_k. As w_k looks random for everyone who does not have the private key you cannot check that there is anything wrong with w_k or w_k xor l_k.
"For simplicity let us assume RDRAND will only attack the Linux RNG."
Assuming the chip is detecting that the Linux RNG is in play is already way out of the realm of simplicity and frankly way beyond what a company like Intel is likely to be able to keep secret given the number of engineers that would have to be aware of this complex functionality.
This whole conspiracy theory hinges on some wild claims that I haven't seen substantiated in the least.
As far as I have followed the discussion there are no hard facts or claims at all, just the general suspicion against a completely closed system of an US company.
I, too, don't believe that Intel adaptively generates its RNG to spoil the Linux RNG. But be reminded that what would have been wild conspiracy theories just half a year ago is now common believe (NSA deliberately introducing vulnerabilities in software and even in cryptographic standards, routinely by-passing TLS).
Given all we know (and don't know) I think it would be prudent to mix Intel's RNG with the other randomness sources using a cryptographically strong primitive and not just XOR. Personally I'm enthusiastic about Keccak as a reseedable RNG, but these modes will probably be standardized no earlier than fall 2014.
> Assuming the chip is detecting that the Linux RNG is in play is already way out of the realm of simplicity
I meant simplicity of my argument. As an answer to this paragraph I argued that it is indeed possible to generate malicious output that appears completely random:
> If this were true and you set up a repeatable test situation in which you force the other parts of the RNG to generate the same numbers prior to RDRAND and then did the RDRAND and captured the results then I don't see how one could argue RDRAND is compromised in this way if the results coming out of it over time even appear to be statistically random.
