Surely the original 'false positive rate' is only wrong if the hash function is 'bad', and thus Bloom wasn't 'incorrect' as much as assuming a perfect hash function.
In the extreme case, imagine that your k hash functions are nearly identical: hash function `H_i` differs from `H_1` only in that the outputs for the `1`st and `i`th elements in the input space are swapped. Of the `N` elements in the input space, all but `k` will completely collide.
If you could assume a perfect hash function, you could treat 64 or 80 bits as secure in a wide swath of inappropriate contexts.
You're confusing statistically uniform with cryptanalytically secure. By that logic you can 'debunk' plain old hash tables because someone might feed you keys that all hash to the same slot.