I think the key is common sense good security engineering with strong verification of security models. We have had a few networked systems that can be target by malicious actors for a while now, like factory control systems; thankfully there's no big motivated enemy out there, but the risk will surely increase in the face of possibility of mass modification (overriding the manufacturer's firmware en masse), even accidental modification by the manufacturers.
Of course, if the user purposely modifies his own control systems he's going to be liable for any damage it causes.
Your faith in "common sense good security engineering" saving the day flies in the face of years of IT security reality. Computer security has been discussed, both formally and informally, since there have been networked computers (have a look at the "Ware Report", prepared by the RAND corporation for the DoD in 1967), yet the general state of IT security is horrendous.
The Toyota uncontrolled acceleration bug and the subsequent software quality issues were reported by Phil Koopman make me think the auto industry isn't particularly qualified to take on security, reliability, or formal verification of their software. (Toyota might be a single example, but I find it hard to believe that software quality is generally any better across the industry. VW's software-based emissions defeat device makes me that much more suspicious of the industry as a whole.)
Attackers are going to be everyone from nation-states who want to surveil their citizenry to individuals who "game" the algorithms in other drivers' autonomous cars to allow themselves to move more quickly through traffic. I'm sure that staged accident insurance fraud, carjacking, hit-and-run, and other vehicle-related crimes will evolve in the face of the rise of autonomous vehicles.
I expect an outcry from law enforcement and "grass roots" groups to demand draconian measures to "lock down" autonomous vehicles. I expect that copyright maximalists and others who would benefit from restrictions on the use of general purpose computers will jump on the bandwagon too. I'm envisioning a future like Stallman's "Right to Read", but applied to our cars and any computers that interface with them.
Sadly, I expect that the fact that general purpose computers are a major component of self-driving cars will mean further legislative restriction on the use of general purpose computers. The idyllic vision of everyone riding in self-driving cars with "tamper-proof" and "trusted" firmware / processors is seductive (and arguably easier to build) than developing systems that are resilient in the face of malicious actors. Cars will end up being computers you ride in, running code that you legally prevented from examining or modifying. I hate the thought of it, but I can't see another scenario actually playing out.