You are "regressing" out some variable (here the batch) from the data, so some values can be negative. For differential analysis, you can include the batch effects in the model (e.g. in DESeq2, ~ batch + condition) , instead of using data that is been corrected for batch effects. The batch corrected data might be useful for visualizations like heatmaps or if you goal is something else than differential analysis (e.g QTL analysis)

This is normal and expected for genes with low counts, you cannot avoid this. combat was developed for arrays, not counts back in the day. Alternatively, use ComBat-Seq from the sva package on the raw counts to remove the batch and then normalized these data. This will preserve the integer-nature of the counts and will not produce negative values.