---
id: 20260822-T0-06
title: "LLM安全对齐存在跨语言漏洞：非英语场景保护失效"
title_en: "Cross-Lingual Safety Gap: LLM Safety Filters Fail in Non-English Languages"
url: https://ai.daily.yangsir.net/daily/20260822-T0-06
issue_date: 2026-08-22
publish_date: 2026-08-21T04:00:00.000Z
category: research
source_name: "arXiv cs.AI"
source_url: https://arxiv.org/abs/2608.18131
---

# LLM安全对齐存在跨语言漏洞：非英语场景保护失效

arXiv 新论文揭示大语言模型安全对齐训练存在严重的英语中心化问题。当安全过滤器在非英语语言中失效时，影响会立即体现在用户端，语音助手和口语对话系统可能产生不当内容。该研究提醒开发者注意多语言模型的安全对齐不足，建议在部署非英语场景时额外验证安全性。

## English Version

**Cross-Lingual Safety Gap: LLM Safety Filters Fail in Non-English Languages**

A new arXiv paper highlights a serious English-centric issue in LLM safety alignment training. When safety filters fail for non-English languages, consequences are immediate: voice assistants and spoken dialogue systems may produce inappropriate content. The research urges developers to verify safety in non-English deployments.

---

**来源**：[arXiv cs.AI](https://arxiv.org/abs/2608.18131)

**详情页**：https://ai.daily.yangsir.net/daily/20260822-T0-06

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*