You absolutely have to use identical input data, hence download the most recent version of the database, if you want to make any claims about reproducibility of the result of two different programs. Also, you didn't indicate which version of local Blast you used, e.g. Blast+, you should use the most recent version here too, and report GIs for hits that potentially got reclassified taxa. Indeed you should look up the GI's of the top hits and compare their annotation.

Also, sequences, due to non-redundancy, might have multiple ID's and also multiple taxa. The sequence is possibly from a viral sequence integrated into the plant genome and therefore annotated with the plant's taxid.

Other misconceptions:

1. Results are not "wrong". A wrong alignment would be one that shows sequences which do not align or that couldn't be reproduced using smith-waterman given the obtained score. You could spot them immediately from the homology string, and therefore this didn't happen.
2. The output of blast, especially looking only at the top hits gives no definitive answer about what the query sequences "are" (I guess you mean which organism it is coming from). In case of multiple hits, the ordering might be arbitrary, some alignments might get identical e-value and score, and therefore you cannot infer the origin of the sequence this way.