The NSA appears to have chosen RC6 for securing the communication channels for its spyware[1].
I have found the choice curious as the cipher has become somewhat obscure after AES was chosen and it happens to be pretty much the only cipher with data-dependent rotations of any renown.
Under the assumption that the NSA may know something we don't, is there a good reason for why data-dependent rotations would harden a cipher? My attempt at answering this and my understanding is that pretty much all currently popular ciphers opt for simplicity and as a consequence may have exploitable mathematical structure (that is currently unknown to be weak). Another point to this hypothesis is that SHA-2 as put forward by the NSA also happens to be a 'messy' unbalanced-Feistel permutation with a lot of interplay between various nonlinear constructs.
Another possibility for why RC6 was chosen is that it may be efficient when AES-NI and SSE/AVX are disallowed (binaries using them would be trivially flagged as engaging in potential encryption). However after consulting some latency and throughput tables for modern hardware it doesn't appear likely - MUL instructions have 3 times the latency of addition (for eg. ARX) and 1/4th the throughput.