Towards Understanding Jailbreak Attacks in LLMs: A Representation Space Analysis

TL;DR

This paper investigates why certain jailbreak attacks succeed in misleading large language models by analyzing their representation space, aiming to understand the intrinsic properties that enable harmful prompts to bypass safety measures.

Contribution

It introduces a novel analysis of LLM representation space to explain jailbreak attack success, proposing a hypothesis about movement towards harmless prompt representations and validating it experimentally.

Findings

Successful attacks move harmful prompts' representations towards harmless prompts.

Incorporating representation space objectives improves attack effectiveness.

Provides insights into LLM understanding of harmfulness information.

Abstract

Large language models (LLMs) are susceptible to a type of attack known as jailbreaking, which misleads LLMs to output harmful contents. Although there are diverse jailbreak attack strategies, there is no unified understanding on why some methods succeed and others fail. This paper explores the behavior of harmful and harmless prompts in the LLM's representation space to investigate the intrinsic properties of successful jailbreak attacks. We hypothesize that successful attacks share some similar properties: They are effective in moving the representation of the harmful prompt towards the direction to the harmless prompts. We leverage hidden representations into the objective of existing jailbreak attacks to move the attacks along the acceptance direction, and conduct experiments to validate the above hypothesis using the proposed objective. We hope this study provides new insights into understanding how LLMs understand harmfulness information.