LLMStinger: Jailbreaking LLMs using RL fine-tuned LLMs

TL;DR

LLMStinger uses reinforcement learning to fine-tune attacker LLMs, automatically generating adversarial suffixes that significantly improve jailbreak attack success rates across multiple large language models.

Contribution

This paper introduces a reinforcement learning-based method for automatically generating adversarial prompts, outperforming existing red-teaming techniques in jailbreak success rates.

Findings

Achieved +57.2% ASR on LLaMA2-7B-chat

Achieved +50.3% ASR on Claude 2

High success rates on GPT-3.5 and Gemma-2B-it

Abstract

We introduce LLMStinger, a novel approach that leverages Large Language Models (LLMs) to automatically generate adversarial suffixes for jailbreak attacks. Unlike traditional methods, which require complex prompt engineering or white-box access, LLMStinger uses a reinforcement learning (RL) loop to fine-tune an attacker LLM, generating new suffixes based on existing attacks for harmful questions from the HarmBench benchmark. Our method significantly outperforms existing red-teaming approaches (we compared against 15 of the latest methods), achieving a +57.2% improvement in Attack Success Rate (ASR) on LLaMA2-7B-chat and a +50.3% ASR increase on Claude 2, both models known for their extensive safety measures. Additionally, we achieved a 94.97% ASR on GPT-3.5 and 99.4% on Gemma-2B-it, demonstrating the robustness and adaptability of LLMStinger across open and closed-source models.