The following expression return NaN in R 2.14.0, but it should not. Furthermore, the same expression seems to have the correct behavior in R 2.10.1
In pbinom(q, size, prob, lower.tail, log.p) : NaNs produced
This was caused by r51327
Well, it does not 'fail' as claimed, it does not return -Inf as before (which is not the actual answer).
The fix in r53127 was far too optimistic in using bpser in a region it is clearly documented not to support: this sort of thing gives other people's code an undeserved bad reputation.
I've added some checks to avoid NaNs (and they do give -Inf in this case).