As my simulation runs, I see that in the .ANI file some of the water molecules simply "run away" from the simulation box
This might be OK. As far as I remember, if you initially give the molecule outside the unit cell, SIESTA will continue to use them and will not translate into the cell during CG.
So this is maybe true for MD as well.Remember that lattice vectors does not mean the boundary of the box, but just the vectors of translation.
