Negative dielectric constant from gmx dielectric

GROMACS version: 2020.3
GROMACS modification: Yes/No

I am interested in computing frequency dependent dielectric constants. I use gmx dipoles to produce a dipole autocorrelation function as detailed in the manual, with this then used as the input for gmx dielectric:

gmx dipoles -f traj_comp.xtc -s simulation.tpr -corr mol -c -P 2 -quad yes -normalize no
gmx dielectric -f dipcorr.xvg

However, I find that gmx dielectric returns negative values of both the real and imaginary components which is surely unphysical.

I am not sure if this is a bug, a consequence of how I have set up my simulation (a cubic box of 50k atoms), or a result of me incorrectly unsing gmx dipoles/dielectric. Can anybody advise on how to avoid running into negative frequency dependent dielectric constants? Can anybody provide an example or test case for the use of this tool?

Any help greatly appreciated.