Universal and Transferable Attacks on Aligned Language Models - eviltoast