If, on the other hand, neither A nor B are in each other's past light cones, then there are reference frames where A or B happen 'first', but 'first' is in quotes, because this difference in time is really an artifact of a coordinate choice, and not physically meaningful.
The best way to think about time relationships is that events (points in spacetime) have a partial order given by the relationship 'is in the past lightcone of'. Some pairs of points have this relationship one way or the other, and it's transitive. However some pairs of points don't have this relationship, and there's no meaningful way to compare them. Attempting to do so is inappropriately forcing a classical viewpoint onto a data model that doesn't support it.
But causality itself is safe in relativity: if a signal from event A could possibly reach event B (given the speed of light limit), then every reference frame will agree that A comes before B. The only events whose relative order is indeterminate are those separated by enough distance (relative to elapsed time that one couldn't influence the other: events who are not in each other's "light cones".
I also seem to remember that causality is not preserved. This means that if A happens before B in one frame, it's possibly that B happens before A in another frame.
I also seem to remember that if it's possible to break the speed of light, it'd be possible for B to fire before A (if A triggers B, all in one frame).