Bug? Numerical Issue? Help! Optimal objective value of relaxation is negative despite being the sum of non-negative variables?

So I was using slack variables (indicating how much a constraint is violated for > 0, = 0 means that the constraint is fulfilled) to find any feasible solution as fast as possible (why doesn't Gurobi have a feature for that btw.?), I looked at the .lp files, the objective function is the sum of those slack_variables (and for some reason a lot of variables with constant factor zero?) and they all have lower bounds >= 0. So neither the objective, nor the relaxation objective should ever be able to be negative. While it works fine for very small instances, for medium- and large-sized instances, I get huge negative bounds (>=10^6). Does anyone have any idea what might be going on? This makes no sense to me...

Best regards,

Tobi