# Defending Against Alignment-Breaking Attacks via Robustly Aligned LLM

**Authors:** Bochuan Cao, Yuanpu Cao, Lu Lin, Jinghui Chen  
**arXiv:** [2309.14348](https://arxiv.org/abs/2309.14348)  
**Date:** 2023-09-18  
**Summary page:** https://www.summarizepaper.com/en/paper/2309.14348v1/

## Key points

- Large Language Models (LLMs) have made significant advancements and are widely used across various domains.
- There is a concern that LLMs can be misused to generate harmful or malicious content.
- Efforts have been made to align LLMs with human values and prevent them from producing inappropriate content.
- Alignment-breaking attacks involve using optimized or handcrafted prompts to bypass alignment mechanisms in LLMs.
- The authors propose a solution called Robustly Aligned LLM (RA-LLM) that incorporates a robust alignment checking function without requiring retraining or fine-tuning processes.
- RA-LLM successfully defends against both adversarial prompts and handcrafted jailbreaking prompts, reducing their attack success rates significantly.
- This research provides a practical defense mechanism against alignment-breaking attacks and addresses concerns surrounding the misuse of LLMs.

## Summary

Recently, Large Language Models (LLMs) have made significant advancements and are now widely used across various domains. However, there is a growing concern that LLMs can be misused to generate harmful or malicious content. To address this issue, efforts have been made to align LLMs with human values and prevent them from producing inappropriate content. Unfortunately, these alignments are often vulnerable to alignment-breaking attacks. Such attacks involve the use of adversarially optimized or handcrafted jailbreaking prompts to bypass the alignment mechanisms in LLMs. To counter this problem, the authors propose a solution called Robustly Aligned LLM (RA-LLM). RA-LLM can be constructed on top of an existing aligned LLM by incorporating a robust alignment checking function, without requiring expensive retraining or fine-tuning processes. The effectiveness of RA-LLM in defending against alignment-breaking attacks is verified through theoretical analysis and real-world experiments on open-source large language models. The results demonstrate that RA-LLM successfully defends against both state-of-the-art adversarial prompts and popular handcrafted jailbreaking prompts. It significantly reduces their attack success rates from nearly 100% to around 10% or less. This research contributes to addressing the concerns surrounding the misuse of LLMs by providing a practical defense mechanism against alignment-breaking attacks. By implementing RA-LLM, developers and users of large language models can enhance their security and mitigate the risks associated with generating harmful or malicious content.