The card design isn't so much security as avoiding scarce real estate on a PC (which, at somewhere over 130 million units fielded, is a not inconsiderable market segment). If this were a dongle device, I'd want it on a parallel port. Many machines don't have a spare serial port, and transparent dongles would be harder to do there, anyway. But transparent parallel port dongle technology is already established.
I agree. I have constructed a parallel port RNG that sampled a blank AM radio band for its source. The data lines give plenty of power to the device, and there are dedicated feedback lines (busy, paper_out, &c). However, I had a very informative discussion with Eric Hughes at CF '94 where I learned that this was the wrong way to go to get good random numbers. Maybe he would like to comment since I don't believe I can do justice to his argument.