Your question reminded me of a post about numerical instability on the
FDS discussion group:
http://groups.google.com/group/fds-smv/browse_thread/thread/7d39ee2c0144acb6In that discussion, Kevin mentions that to simulate a broken window, it may work better to include a buffer region between the outer bounds of your building and the bounds of the computational domain. My impression is that the trick is to have some transfer (leakage?) between the inside and the outside as opposed to suddenly opening an airtight container.