This was in my opinion one of the best talks at defcon a few years back. I replicated this with some CDN domains, and was amazed at the number of bits that were being flipped out there that drove traffic to my test systems. It would have been very easy to serve up content that could have caused a lot of problems.

It would be really interesting to see what mitigation would be possible if this became a serious problem. I suppose the simplest solution is for the responsible cdns to just buy up all of their bitneighbors.

Or shifting most devices to using some form of ECC memory

I wonder how many "single bit errors" are due to hardware faults vs. how many are due to software accidentally setting/unsetting a single bit binary flag through a wild pointer. It would also be interesting to know what percentage of the errors were a zero-to-one flip vs. a one-to-zero, and how well this matches with the expectation of the direction of actual single-bit hardware errors (which I also don't know). Actually, it's also interesting to know which bits were more likely to change; presumably hardware errors are distributed evenly over the eight different bits of a byte, while one might guess that flag bits erroneously (re)set by software might be more likely to be in the low end of a byte.