It was my bug and a change in the driver exposed it. I was freeing some CUDA host memory in a thread without setting the correct CUDA context. Previously it apparently didn't matter. One might think that the address is sufficient, but it is not now true.
Still, nVidia should return a proper error and not crash.