I'm doing survival analysis on several treatment groups. When I select all the groups for comparison, the results are not significant. But when I select two groups, I sometimes get a significant result, sometimes not, depending on which two I select. Is this correct?

This is the "multiple comparisons" problem. When you compare all groups--are they the same?-- you are asking a different question than when you pick out two and ask the same question. The latter, used repeatedly, is really not appropriate unless you make an adjustment for multiple comparisons, something Prism doesn't help you with. A simple approach is to multiply all the P values by the number of comparisons you make, and still use 0.05 as the cutoff.