---
id: 20260623-T0-03
title: "研究发现：提示词注入本质是角色混淆，系统指令可被伪装"
title_en: "Prompt Injection Explained as Role Confusion: System Instructions Can Be Spoofed"
url: https://ai.daily.yangsir.net/daily/20260623-T0-03
issue_date: 2026-06-23
publish_date: 2026-06-22T23:59:53.000Z
category: insight
source_name: "Simon Willison"
source_url: https://simonwillison.net/2026/Jun/22/prompt-injection-as-role-confusion/#atom-everything
---

# 研究发现：提示词注入本质是角色混淆，系统指令可被伪装

Simon Willison撰文解析一项关于提示词注入的研究，指出该类攻击的核心在于AI模型无法区分用户输入和开发者指令。文章将复杂的学术发现转化为通俗的博客形式，解释了当系统提示词与用户数据混合时，模型容易产生角色混淆。这种机制允许攻击者通过伪装数据覆盖系统指令，从而绕过安全限制。

## English Version

**Prompt Injection Explained as Role Confusion: System Instructions Can Be Spoofed**

Simon Willison provides a blog-style breakdown of a new paper characterizing prompt injection as role confusion. The research highlights that LLMs struggle to distinguish between user input and developer instructions. When system prompts and user data are mixed, models experience role confusion, allowing attackers to override safety instructions by spoofing data.

---

**来源**：[Simon Willison](https://simonwillison.net/2026/Jun/22/prompt-injection-as-role-confusion/#atom-everything)

**详情页**：https://ai.daily.yangsir.net/daily/20260623-T0-03

---

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