Implementation details will change with the next release, to use an updated version of the Baiillie-PSW method.
https://arxiv.org/abs/2006.14425
This of course remains probabilistic. Here is Robert Baillie's remark though, which strikes me as quite apt for the problem at hand. "There are zero known counterexamples to the original BPSW, and we expect the new test to have even fewer :)"