I've experienced a segfault with mdrun.
The crash is reproduced with both 4.5.4 release and git master, no matter whether single-threaded or multi-threaded, and whether single or double precision.
The cause of the crash seems to be unstable box size on NPT sampling. (see md.log).
Although this is due to blowing-up simulation, I guess it is better to stop rather than crash (to prevent silent degradation).
Way to reproduce:
grompp_d -c conf.g96 mdrun_d
Around 1300 steps the program crashes.