I've had a minimed paradigm for about 8 years now, and all of what Scott said makes sense. In addition, there are a few more things which make this impractical. I assume the researcher is trying to hack the "Remote" option. Not only do you need to turn the remote option on, you need to add IDs of the remotes to the pump itself. So unless you can figure out how to add IDs remotely, you have to find someone with a remote, and get the ID from the remote.
Second, there's a limit (at least on my Paradigm version) of 20 units of insulin at a time. I haven't tried this, but I think there's a system to prevent you from giving multiple 20 unit boluses at a time. Since I take around 14 units for some meals, 20 units of insulin is conceivable to overcome just by eating sweets, and there's always glucagon injections in a pinch. My pump makes a sound when it is done giving a bolus, meaning the diabetic could notice that a bolus was given (perhaps the beep is turned off for continuous glucose monitoring systems though).
Finally, hypoglycemia is rarely fatal. From wikipedia: "In nearly all cases, hypoglycemia that is severe enough to cause seizures or unconsciousness can be reversed without obvious harm to the brain." So even if you figure out how to give a remote bolus and succeed, it isn't likely to kill the diabetic.
According to this 5000 respondent survey the failure rate is 54.2%, but the article points out that over 30 million consoles have been sold. I would place little confidence in the 5000 person survey.
Actually, with a population of 30 million, you can be 99% confident of the result with a confidence interval of +-2% with a sample size of 4,160. Check these numbers here. This means you know with 99% confidence that the actual population failure rate is between 52.2% and 56.2%. Sample sizes don't need to be as large as most people think to produce statistically significant results. Of course, that calculation assumes a random sample from the population, whereas this was sampled only from readers of Game Informer. I could see an argument that the numbers are skewed by selection bias, but the sample size is large enough.
Another megabytes the dust.