While I get the humor in your comment, I'd be curious if 'relaxing' some of the behavior parameters would result in some not-very-useful conclusions from testing:

"Tests indicate most efficient driving technique to include lots of honking, flashing high beams, and spraying winshield wiper fluid onto cars following too close on the highway."

In situations which usually induce anger in people on the roads, it seems more likely that any resulting behaviours resembling road rage only disadvantage the driver, if the goal is safety.

So if anything, wouldn't any sort of sufficiently global optimisation only veer the ai away from behaviours resembling road rage over time?

how do you define safety programmatically so the vehicle knows what's considered 'safe'? Measurement of g-force in sudden deceleration, acceleration, lateral movement? Heart beat per minute of passengers?

Most passenger cars are far more capable than average drivers can exploit, so there is quite a bit of room for cars to drive very aggressively from human's perspective, but completely 'safe' from machine's perspective.

Going one step further, some AI will be designed to be aggressive; for example, law enforcement, military, security escorts, etc. -- I'm guessing that will it is very likely that you're right that some consumer AI will be aggressive, if it become too common, it'll be shut down.