Ask HN: How much negative testing is done by Tesla and other self-driving teams?

In software testing, there's positive testing to check that software will perform desired behavior, and negative testing to ensure that the software will avoid undesired behaviors.

We know that self-driving cars have many thousands of hours of positive testing done. But what I'm wondering is how much adversarial testing is done?

For example, to test that the car doesn't get misled by old lane markers when a road is under construction, wouldn't you at least do millions of simulated tests with randomly placed painted lines overlaid on the virtual road?

And to test that the car doesn't run into stopped emergency vehicles or tractor trailers crossing its path, wouldn't you at least do thousands of tests on a road containing stopped vehicles?

Not to mention if someone accidentally guns the accelerator while approaching a sharp curve. Shouldn't the car know that it's approaching the curve and issue a warning?

Does anyone attempt to do the driving equivalent of fuzzing to self-driving cars?