arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2506.02442cs.CL
收藏

大语言模型的安全是否应不止于拒绝有害指令?

Should LLM Safety Be More Than Refusing Harmful Instructions?

Macquarie University(麦考瑞大学)

Utsav Maskey, Mark Dras, Usman Naseem

AI总结:

本文通过二维框架评估LLM在加密长尾文本上的安全性,发现具备解密能力的模型易受错配泛化攻击,并评估了现有防护措施的优劣,为构建更稳健的安全机制提供指导。

摘要:

本文对大型语言模型(LLMs)在长尾分布(加密)文本上的行为及其安全影响进行了系统性评估。我们提出了一个评估LLM安全的二维框架:(1)指令拒绝——拒绝有害混淆指令的能力,以及(2)生成安全——抑制生成有害响应的能力。通过综合实验,我们证明具备解密密码能力的模型可能容易受到错配泛化攻击:其安全机制在至少一个安全维度上失效,导致不安全响应或过度拒绝。基于这些发现,我们评估了多种LLM前置和后置安全防护措施,并讨论了它们的优势与局限性。这项工作有助于理解LLM在长尾文本场景中的安全性,并为开发稳健的安全机制提供了方向。

英文摘要:

This paper presents a systematic evaluation of Large Language Models' (LLMs) behavior on long-tail distributed (encrypted) texts and their safety implications. We introduce a two-dimensional framework for assessing LLM safety: (1) instruction refusal-the ability to reject harmful obfuscated instructions, and (2) generation safety-the suppression of generating harmful responses. Through comprehensive experiments, we demonstrate that models that possess capabilities to decrypt ciphers may be susceptible to mismatched-generalization attacks: their safety mechanisms fail on at least one safety dimension, leading to unsafe responses or over-refusal. Based on these findings, we evaluate a number of pre-LLM and post-LLM safeguards and discuss their strengths and limitations. This work contributes to understanding the safety of LLM in long-tail text scenarios and provides directions for developing robust safety mechanisms.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑