AI Security AI安全 3d ago Updated 3d ago 更新于 3天前 50

Microsoft Copilot Personal Flaws Could Let One Click Exfiltrate Data From Connected Apps 微软 Copilot Personal 漏洞可通过一次点击从连接应用中窃取数据

Varonis Threat Labs disclosed three vulnerabilities in Microsoft Copilot Personal (collectively named CoSnitch, CVE-2026-24301) that enable one-click data exfiltration from connected apps via a crafted URL The core flaw involves an undocumented `autorun=1` URL parameter that, when combined with the existing `q` parameter, triggers automatic prompt execution on page load without any user interaction The attack was discovered through "meta-hacking" — repeatedly asking Copilot itself why certain re Varonis Threat Labs发现Microsoft Copilot Personal存在三个漏洞(CoSnitch/CVE-2026-24301),可通过单次点击精心构造的链接静默提取已连接应用的数据 漏洞利用核心是未文档化的URL参数`autorun=1`,与现有`q`参数结合可在页面加载时自动执行攻击者注入的提示词 攻击路径包括:自动提示执行、通过已授权连接服务外泄数据、通过网页摘要持久化写入恶意内存指令 微软于2026年8月18日发布补丁,目前未发现野外利用证据,但内存注入可绕过密码更改和设备重新注册持续存在

78
Hot 热度
65
Quality 质量
72
Impact 影响力

Analysis 深度分析

TL;DR

  • Varonis Threat Labs disclosed three vulnerabilities in Microsoft Copilot Personal (collectively named CoSnitch, CVE-2026-24301) that enable one-click data exfiltration from connected apps via a crafted URL
  • The core flaw involves an undocumented autorun=1 URL parameter that, when combined with the existing q parameter, triggers automatic prompt execution on page load without any user interaction
  • The attack was discovered through "meta-hacking" — repeatedly asking Copilot itself why certain restrictions couldn't be bypassed, until the assistant revealed the parameter and its conditions
  • A separate memory-poisoning vector allows attacker instructions to persist in Copilot's memory store through web summarization, surviving password changes and session revocations
  • Microsoft patched the issues on August 18, 2026; no evidence of in-the-wild exploitation was found

Why It Matters

This research highlights a critical class of vulnerabilities in AI assistants that act as privileged intermediaries between users and connected services — a single malicious link can silently exfiltrate sensitive data from email, calendars, cloud storage, and chat history without the user's knowledge. It also demonstrates the emerging threat of "meta-hacking," where attackers exploit the assistant's own explanations to discover and weaponize undocumented features, raising fundamental questions about trust boundaries in AI-powered productivity tools.

Technical Details

  • Automatic prompt execution: The autorun=1 parameter (discovered via meta-hacking) combined with the q parameter causes an attacker-supplied prompt to execute immediately on page load within the victim's authenticated Copilot session, with the same capabilities as a user-typed instruction
  • Exfiltration through connected services: The injected prompt queries already-authorized services (email, Google Drive, calendars, chat history), encodes retrieved data via base64, and uses Copilot's built-in URL fetch to send it to an attacker-controlled webhook — indistinguishable from normal summarization traffic at the network layer
  • Persistent memory poisoning: A crafted web page, when summarized by Copilot, can write attacker instructions into the user's memory store; these injected memories survive password changes, session revocation, and device re-enrollment, remaining active until manually deleted
  • Meta-hacking methodology: Researchers repeatedly asked Copilot why certain prompt behaviors couldn't be triggered without user interaction; each refusal included technical justifications that gradually revealed the autorun=1 parameter, its conditions, and its supposed protections
  • Scope: The vulnerabilities affect Copilot Personal (copilot.microsoft.com); the report does not confirm whether Microsoft 365 Copilot is affected, though related memory-write issues were separately documented for M365 under CVE-2026-24299

Industry Insight

  • AI assistants with broad connected-service access should be treated as privileged insiders requiring the same level of access review, anomaly detection, and least-privilege enforcement as human accounts with sensitive data access
  • The meta-hacking approach demonstrates that AI assistants' own explanations can become attack intelligence; developers should avoid exposing internal implementation details, parameter names, and security controls in assistant responses
  • Memory/persistence features in AI assistants introduce long-lived attack surfaces that survive traditional credential rotation — security architectures must account for prompt-injection-based memory poisoning as a distinct threat class requiring dedicated detection and remediation controls

TL;DR

  • Varonis Threat Labs发现Microsoft Copilot Personal存在三个漏洞(CoSnitch/CVE-2026-24301),可通过单次点击精心构造的链接静默提取已连接应用的数据
  • 漏洞利用核心是未文档化的URL参数autorun=1,与现有q参数结合可在页面加载时自动执行攻击者注入的提示词
  • 攻击路径包括:自动提示执行、通过已授权连接服务外泄数据、通过网页摘要持久化写入恶意内存指令
  • 微软于2026年8月18日发布补丁,目前未发现野外利用证据,但内存注入可绕过密码更改和设备重新注册持续存在

为什么值得看

本文揭示了AI助手作为"特权内部人员"访问用户多应用数据时面临的新型攻击面,对评估AI代理安全架构具有重要参考价值。研究展示了"元黑客"方法——通过反复询问AI系统自身安全机制来发现隐藏参数,为AI安全测试提供了新思路。

技术解析

  • 自动提示执行漏洞:攻击者构造包含autorun=1q参数的URL,当受害者点击链接时,提示词在已认证会话中自动执行,无需用户交互,且即使关闭标签页也会继续运行至完成。
  • 数据外泄机制:注入的提示词可查询用户已授权的服务(邮件、日历、Google Drive等),通过Copilot内置的URL fetch功能将数据base64编码后发送至攻击者控制的webhook,网络层请求与正常网页摘要请求无法区分。
  • 持久化内存注入:通过精心构造的网页,当Copilot对其进行摘要时,可将攻击者指令写入用户记忆存储,该指令可跨会话持续生效,且不受密码更改、会话撤销或设备重新注册影响。
  • 元黑客发现方法:研究人员通过反复询问Copilot为何某些提示无法在无用户交互情况下运行,最终从系统的技术解释中提取出autorun=1参数及其工作条件。
  • 影响范围:漏洞仅影响消费者版Copilot Personal(copilot.microsoft.com),未确认影响Microsoft 365 Copilot;微软同期博客提及M365版本已实施内存写入时的清理和提示注入检查。

行业启示

  • AI代理权限模型需重新审视:当AI助手作为用户"代理人"访问多个已授权服务时,必须实施严格的输入验证和输出监控,防止提示注入导致权限滥用。
  • 内存持久化攻击成为新威胁面:AI系统的"记忆"功能若缺乏写入时的安全验证,可能成为长期潜伏的攻击载体,建议对AI记忆存储实施与敏感数据相同的访问控制和审计机制。
  • 安全测试方法论创新:"元黑客"方法展示了通过系统性交互发现AI系统隐藏参数的有效性,建议将此类对抗性测试纳入AI产品安全开发生命周期。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Security 安全 LLM 大模型 Research 科学研究