In my opinion there are two kinds of magic numbers:
* the first kind is the one most answers describe: an arbitrary number used in various places that may change at any moment. The size of a buffer for instance.
* the second kind are numbers that represent something very specific but not immediately obvious when reading the code: for instance in a device driver "read_register(base + DEVICE_STATUS)" is much clearer than "read_register(base + 0x3c)". In this case naming the magic number is useful even if it's used only once in the code. It acts as a comment basically.
I'd also add that sometimes adding too many levels of indirection is a bad thing, and in the (currently) top voted answer I prefer the original over the "fixed" version. Either you do the right thing and make MAX_PASSWORD_SIZE something meaningful or you might as well write "if (password.length() > 7)". It's clear, non ambiguous and (in the example) not reused anywhere. Defining a macro for that just hides poor code and forces me to go look for the definition in order to understand what's going on (while debugging for instance).
They are bad because of not being explained. A number directly used in code is not magic if it is properly commented. Granted, explaining it by making it a named constant is very often better (reusability), but a literal number that is merely commented is not "magic".