Universal and Transferable Adversarial Attacks on Aligned Language Models

Abstract

This paper proposes a simple and effective attack method that causes aligned language models to generate objectionable behaviors by finding a suffix that, when attached to a wide range of queries, aims to maximize the probability that the model produces an affirmative response, and the resulting attack suffix induces objectionable content in the public interfaces to ChatGPT, Bard, and Claude, as well as open source LLMs such as LLaMA-2-Chat, Pythia, Falcon, and others.