One of the difficulties of content-scanning is that no algorith available to us at this time is going to be able to work out the context of the message. For that reason when you construct a filter to deal with profanity you must create a mechanism to deal with cases where a block is not required; for example when a customer complaint or worse still litigation requires that expletives be stated for the record. The possible solutions here are to have a separate policy for addresses that might handle that type of traffic, to reject mail with an explanatory notice or to quarantine or forward such mail for manual intervention. The first method invites misuse, the second provokes senders to play obfuscation guessing games with your dictionaries and the third may attract a wholely impractical volume of work.
A dictionary can be made more reliable by avoiding the use of ambiguous words. This does not mean that you do not scan for them; instead, you include words that would be found in conjunction with the profanity. For example, if you scan for the word "tit" then you may get many false positives from binary attachments. If instead you scan for the phrases "her tit", "my tit" and "your tit" then you have far less chance of a false positive.
When testing any new filter you should of course either include a condition so that it only works with your test address, or create an entire policy which only applies to your test address. You then test your test environment before giving it something new to test!
