The probability can be dialed depending on your needs and you can make it as low as you want.
For a Bloom filter storing 50k hashes with one in quadrillion (10^-15) probability of false positive you need 438kB bitmap.
I would worry more about ability to create software with hashes that exploit false positives in the filter. I guess there is multiple ways to remediate this: the filter could be used only as fallback if the primary service doesn't work. And it could also be rotated regularly. It could be salted (ie. thousands of versions of the filter available) making it difficult to predict which one is going to be used. These are just couple silly ideas off the top of my head.
For a Bloom filter storing 50k hashes with one in quadrillion (10^-15) probability of false positive you need 438kB bitmap.
I would worry more about ability to create software with hashes that exploit false positives in the filter. I guess there is multiple ways to remediate this: the filter could be used only as fallback if the primary service doesn't work. And it could also be rotated regularly. It could be salted (ie. thousands of versions of the filter available) making it difficult to predict which one is going to be used. These are just couple silly ideas off the top of my head.