There are no doubt doses that are lethal even with a toxin filter going full tilt. The filter is unlikely to be a complete molecular sieve (since it is hard to do without blocking the bloodstream). It is more likely specialized tissues that pick up bad molecules and sequesters them. This means that it cannot remove 100% of anything on the first pass through the blood, and hence if you inhale or are injected with something nasty it will have a few circulatory cycles to act. Worse, once it is inside cells it can do its dirty job unhindered. And the filter might get overwhelmed by a large dose. Here is a math model: the filter has receptors that bind toxins, and once bound removes them. This can be modelled by Michaelis–Menten kinetics. I would assume the filter to be very efficient, so it has a low K_m, perhaps 10^-6 and let's say k_cat=10^6 - this is better than what normal enzymes can do, but we are doing nanotech here. So for a heavy dose of toxin the removal rate will be close to 10^6*[receptor] molecules per second. Now, there has to be enough receptors to handle a load of toxin molecules in a single pass of 30 seconds. A breath is about a litre, and if it contains a deadly concentration of 300 mg/m^3 of hydrogen cyanide, it will have 0.3 mg cyanide. That is 6.7e21 molecules. So there has to be 2.2e14 receptors. But now imagine that you double the air concentration. Each breath will add twice as many molecules, and while the first half are all being bound and digested the other ones, a concentration equal to the lethal level, are free to play havoc. Same thing with any other toxin: there is some concentration where every receptor is busy, and any higher concentration will leave residual toxin to do damage. Of course, the real question is how to make receptors that can both pick up cyanide ions and complex proteins like botulin toxin.