In a way, humans are also machines, and we have very similar attack vectors, so just extending our current solutions might be good enough to accommodate self-driving cars on our roads.
I agree it is a very rare event, and will always continue to be. But it would need to be handled, and handled in ways other than, "Hi, here's a 100 car pile-up."
Similarly, and probably less rarely, you'd need to handle scenarios in which another actor wasn't malicious per se, but faulty to a degree that it was functionally malicious. Of course, we've always had to handle the idea that somebody might be driving a car with bad brakes, and that's dangerous enough, but before now a car with bad brakes wouldn't also make the brakes bad for ten cars around them. I think that's worthy of some contemplation.
We'll have to sort out a lot of those scenarios, many by means of "testing in production", but as long as self-driving cars can outperform humans on average, we'll be better off.
The most that can happen is each car will probably end up sending Google certain diagnostic information, but actual environmental sensing will have to be done by the car itself, with maybe some help from aerial cameras or satellites that Google controls.