the crypto scheme was randomized between 3 algos to reduce the statistical data size for each. one was AES, one was alphacrypt, another one was something else. the json protocol exchanged an algo_id which was stored with each message.
as for home-brew, i used the encrypt/decrypt code i found at the provided location. i'm no crypto expert and the guy claims to have a mathematics PhD with "Eleven years of publishing scientific and technical papers (computer science and higher-level mathematics)", so hopefully his security assertions are not entirely without merit.
the key is not stored anywhere. it must be exchanged by other secure means, just like any symmetric encryption/AES.
if you encrypt "a" using a key "foo" and the lib selects a random seed char to tack onto every word, there can (from what i understand) be 255 variations to encode the same plaintext word...and this set varies for each key.
storing the seed, like storing a salt doesn't seem like it would help much. so it's not immediately obvious that having access to many messages encrypted using the same key would allow you to do any kind of trivial statistical analysis other than on word length alone.
if it is in fact trivially breakable, i would love to see it implemented by simply having access to the encryption code and ciphertext of 5 different messages of several words in length encoded using the same key 100 times each.
it'd be awesome to learn more about crypto....and about 100 other topics, too :)
