Jailbreaking as a Reward Misspecification Problem

TL;DR

This paper links LLM vulnerabilities to reward misspecification, introduces ReGap and ReMiss for detecting and exploiting this flaw, and demonstrates their effectiveness in improving adversarial prompt generation and model robustness evaluation.

Contribution

It presents the ReGap metric for quantifying reward misspecification and introduces ReMiss, a novel automated red teaming system that outperforms existing methods in adversarial prompt generation.

Findings

ReGap effectively detects harmful backdoor prompts.

ReMiss achieves state-of-the-art attack success rates.

Attacks transfer well to closed-source models and out-of-distribution tasks.

Abstract

The widespread adoption of large language models (LLMs) has raised concerns about their safety and reliability, particularly regarding their vulnerability to adversarial attacks. In this paper, we propose a novel perspective that attributes this vulnerability to reward misspecification during the alignment process. This misspecification occurs when the reward function fails to accurately capture the intended behavior, leading to misaligned model outputs. We introduce a metric ReGap to quantify the extent of reward misspecification and demonstrate its effectiveness and robustness in detecting harmful backdoor prompts. Building upon these insights, we present ReMiss, a system for automated red teaming that generates adversarial prompts in a reward-misspecified space. ReMiss achieves state-of-the-art attack success rates on the AdvBench benchmark against various target aligned LLMs while preserving the human readability of the generated prompts. Furthermore, these attacks on open-source models demonstrate high transferability to closed-source models like GPT-4o and out-of-distribution tasks from HarmBench. Detailed analysis highlights the unique advantages of the proposed reward misspecification objective compared to previous methods, offering new insights for improving LLM safety and robustness.