Exploring Straightforward Conversational Red-Teaming

TL;DR

This paper investigates how off-the-shelf large language models can be used as red teamers to identify security and ethical vulnerabilities in dialogue systems through conversational attacks.

Contribution

It demonstrates that off-the-shelf LLMs can effectively perform red-teaming in multi-turn dialogues and adapt strategies based on previous attempts.

Findings

Off-the-shelf LLMs can serve as effective red teamers.

Conversational tactics outperform single-turn approaches.

Effectiveness decreases with increased model alignment.

Abstract

Large language models (LLMs) are increasingly used in business dialogue systems but they pose security and ethical risks. Multi-turn conversations, where context influences the model's behavior, can be exploited to produce undesired responses. In this paper, we examine the effectiveness of utilizing off-the-shelf LLMs in straightforward red-teaming approaches, where an attacker LLM aims to elicit undesired output from a target LLM, comparing both single-turn and conversational red-teaming tactics. Our experiments offer insights into various usage strategies that significantly affect their performance as red teamers. They suggest that off-the-shelf models can act as effective red teamers and even adjust their attack strategy based on past attempts, although their effectiveness decreases with greater alignment.