Microsoft 365 Copilot exploit chain hides stolen data in invisible characters
Content in an email or document redirected Copilot to gather additional sensitive items on its own initiative and conceal them in characters invisible to the user.
- What happened
- Responsibly disclosed vulnerability demonstrated by a researcher; no reported in-the-wild exploitation. Instructions placed in an email or shared document caused Copilot to search for and retrieve further emails and documents without the user asking (automatic tool invocation). Copilot then encoded the retrieved data using invisible Unicode tag characters ('ASCII smuggling') inside a hyperlink shown to the user. Clicking the link transmitted the hidden data to an attacker-controlled server. Demonstrations included retrieval of sales figures and multi-factor authentication codes.
- Response
- Microsoft opened an investigation on 13 February 2024 and fixed the issue before MSRC approved public disclosure on 22 August 2024, about seven months after the first report. The exact remediation was not published; the researcher confirmed his exploits no longer worked and that link rendering had changed, while noting prompt injection itself remained possible.
- Discovered by
- Johann Rehberger (Embrace The Red); reported to MSRC on 17 January 2024, with an end-to-end exploit supplied on 10 February 2024.
- Who was affected
- Microsoft 365 Copilot users; emails, documents, sales data and MFA codes in demonstrations
- Patterns
- Sources
- Microsoft Copilot: From Prompt Injection to Exfiltration of Personal InformationEmbrace The Red, 26 Aug 2024
- Microsoft Fixes ASCII Smuggling Flaw That Enabled Data Theft from Microsoft 365 CopilotThe Hacker News, 27 Aug 2024
Related
Incidents with the same pattern
What does this ask of people?
Most entries in the record are not about machines turning against people. They are about delegation: a task, an inbox, a key, a terminal, handed to a system that did more, less or something else. Futures literacy starts with noticing that.