I'm not sure what the bar mitvah attack is exactly, that link is kind of vague.

This paper is based on the statistical biases of pairs of bytes in the output stream of RC4, which they credit originally to Fluhrer and McGrew in this paper from 2000: http://www.mindspring.com/~dmcgrew/rc4-03.pdf

Their main contribution is to combine reasonable guesses about the nature of the plaintext with the biases in the cyphertext to improve the time to recover. They also extend the state of knowledge about the byte pair biases in the early keystream.