In short, all of your reasons are possible, including maybe due sequencing errors. But if your depth is high enough you can be little bit confident in calling them as true somatic events. You should also check ExAC for rare SNPs.

What is the coverage? For example, if it's 20X, then 5% is just 1 read, so probably not real.

Did you visually inspect the mutations in a genome browser like IGV? Do the reads and variants look clean? Anything visual is hard to explain, but after you see a few high-quality variants, it should be relatively easy to spot bad calls.

Ultimately, it's not really possible to tell if the mutation is real. You really need to validate by sequencing the samples again. It's very common that even high quality mutations that get called by multiple variant callers and look very good by visual inspection do not validate.