"I held back a small portion of data for calculating the AUC so I have both training and test data."

You likely performed the whole procedure several times (at least 9 times perhaps)? If so you could average the resulting test scores, TSS in your case, even if they are based on a small sample. That being said, computing evaluation measures on a sample of three sounds rather suspicious and i am wondering whether you held back too little data.

Basically, you can use k-fold cross validation where the skill is computed on the withheld part for the k replicates trained (using k minus one parts) and then averaged.

Or you can use other approaches like subsampling or bootstrapping. Either way one fold or the test percentage shouldnt be too small.

If you wanted to use real independent test data you cannot get away with just three observations but need to sacrifice more.

Finally, TSS is prevalence dependent and in combination with small sample size and background this can be problematic.

You can use the training data to obtain predicted values, and calculate the sensitivity, specificity and then the TSS. Now you can say you have a TSS based on training data. For the testing data, you have to do cross validation or other resampling techniques (as Rainer Ferdinand Wunderlich mentioned) to obtain TSS, since three points cannot support a robust results. In short, you can clearly say your TSS is based on what data you used.

Yeah I'm not exactly sure why I have 3-4 "test" data points. I have crossvalidate selected in MaxEnt as the replicated run type. Unless that is something different.

it may be the result of choosing too many folds. k commonly stays between 3 and 7. (a few) more folds (could) only make sense if you have many hundreds or thousands of presences.