[schoen]

I think there was some research about this that did find some flawed implementations leading to semiprime p or q, but basically you can choose the probability that a probable prime is actually composite to be low enough that the expected number of weak keys ever generated for this reason is still less than one.

(I haven't personally understood the probabilistic primality tests well enough to understand how they guarantee that the probability of error contributed by each round of the test is completely independent of every other round.)