I recently came across this manuscript from the Weissman (Stanford) group, in with the authors describe a problem on the more recent Illumia platforms (HiSeq 3000/4000/X), which can lead to contaminations between multiplexed samples.
It’s something that needs to be corrected for in any analysis investigating public sequences. Usually a read count filter (eg, >3 raw reads) is sufficient to catch indexing errors, but low yield samples (eg, tissue, rare sorts, etc) can be trickier.