Martin quoted Glynn:
OTOH, existing implementations (at least GHC and Hugs) currently read and write "8-bit binary", i.e. characters 0-255 get read and written "as-is" and anything else breaks, and changing that would probably break a fair amount of existing code.
The binary library I posted to the libraries list: http://haskell.org/pipermail/libraries/2003-June/001227.html which is for GHC, does this properly. All characters are encoded using a standard encoding for unsigned integers, which uses the bottom 7 bits of each character as data, and the top bit to signal that the encoding is not yet complete. Characters 0-127 (which include the standard ASCII ones) get encoded as themselves. This is probably not nearly as efficient as encoding characters as themselves, but it's nice to be Unicode-proof ...