什么是提示注入？最关键的 AI 漏洞解析

In 2003, SQL injection emerged as the defining vulnerability of the web era, exploiting a system's inability to differentiate between code and data. By 2025, prompt injection has assumed a similar critical role in AI applications. A vulnerability in GitHub Copilot (CVE-2025-53773), patched in August 2025, allowed attackers to achieve full remote code execution by embedding malicious instructions in a README file. Additionally, Slack AI was exploited to silently exfiltrate API keys from private channels through indirect prompt injection. The architecture of large language models makes this class of attack fundamentally challenging to eliminate. This article explains how prompt injection works, why conventional fixes are ineffective, and what testing for it entails.