I have 90 (3 groups of 30) of precious samples which have huge differences in read depth (ranging from 2 million - 60 million aligned reads). I ran an Anova to check that there wasn't a difference for one particular group. Am I able to run the default normalization for these samples in DESEQ2? Is there some kind of 'sensible' filter to discard the samples that may be skewing normalization??

It should be fine if there isn’t confounding with the condition of interest. The biggest problem is that zeros cannot be scaled, so with confounding of sequencing depth with condition you still have zeros vs positive counts for many genes.