The big concern I have is how reliable this is on virtual machines. Just about all physical machines I'd want to use have high-quality, trustworthy-within-my-threat-model (i.e., "if the NSA wanted to attack my silicon, there's easier silicon for them to attack") hardware random number generators, and all physical machines I'd want to use pick up sufficient randomness from the kernel's entropy magic thing. But virtual machines often don't have access to the hardware RNG, and they don't have access to enough other hardware to populate entropy. It seems like this technique would be particularly risky there ... although I don't think anyone's published an attack on DakaRand yet, so maybe it's fine!
See also Kaminsky's implementation of the same approach in pure JS: https://gist.github.com/PaulCapestany/6148566
and Ryan Finnie's implementation in Perl: http://www.finnie.org/2012/08/14/twuewand-2-0-released/
The big concern I have is how reliable this is on virtual machines. Just about all physical machines I'd want to use have high-quality, trustworthy-within-my-threat-model (i.e., "if the NSA wanted to attack my silicon, there's easier silicon for them to attack") hardware random number generators, and all physical machines I'd want to use pick up sufficient randomness from the kernel's entropy magic thing. But virtual machines often don't have access to the hardware RNG, and they don't have access to enough other hardware to populate entropy. It seems like this technique would be particularly risky there ... although I don't think anyone's published an attack on DakaRand yet, so maybe it's fine!