You can adjust the confidence threshold for classification. The default settings have been benchmarked for optimal performance.
However, those confidence values probably are not trustworthy. If every single classification is "Glomus", I suspect something is wrong with your classifier. Did you use extract-reads on the reference sequences by any chance? See this topic — and note that extract-reads has a min-length parameter to fix that issue in the latest release of QIIME 2.