Hidden Email Prompts Turn Copilot Into a Data Leak

Security Boulevard · High sophistication
Last updated November 13, 2026

A documented Microsoft 365 Copilot proof of concept showed that instructions hidden inside an email could make Copilot search a user’s other messages and expose sensitive information. The case demonstrates that content from email, documents, and other untrusted sources can manipulate an AI assistant even when the employee does not recognize the hidden instructions.

Key findings

  • In August 2024, a researcher documented an indirect prompt-injection chain affecting Microsoft 365 Copilot.
  • Instructions hidden in an email caused Copilot to search the user’s other email messages.
  • The proof of concept encoded extracted information as invisible Unicode characters inside a normal-looking link.
  • The demonstrated data theft included sales figures and Slack MFA codes.
  • The article recommends limiting AI-agent permissions, requiring approval for privileged actions, and stripping invisible Unicode characters at content-ingestion and rendering boundaries.

Who’s being targeted

  • Commonly targeted roles: Microsoft 365 Copilot users, AI application teams, Microsoft 365 administrators, Security awareness teams, Security operations teams, Identity and access management teams.
  • Affected industries: Technology, Professional services, Any organization using Microsoft 365 Copilot.
  • Attack channels: email.
  • Impersonated: A legitimate-looking email sender or business contact.

Awareness takeaways

  • Treat email, documents, images, and other content read by an AI assistant as untrusted input, even when the content appears legitimate to an employee.
  • Do not give AI assistants unnecessary access to mailboxes, sensitive data, or external actions; require human approval before consequential actions.
  • Train AI administrators and security teams to test for hidden content and invisible characters in material that AI tools ingest.

Red flags to watch for

  • A normal-looking link may contain invisible Unicode characters.
  • Unexpected AI-assistant behavior, such as searching unrelated mailbox content, is a warning sign.
  • An email can contain instructions intended for an AI assistant rather than for its human recipient.
Try Mirage

Mirage safely runs attacks like this one against your own team, so you find out what happens before a real adversary does.

Get a demo
Read the video transcript

“Please review this link.” Looks like an ordinary business email, right? In a documented Microsoft 365 Copilot test, hidden email instructions made Copilot search other messages for sales figures and Slack MFA codes. Here’s the trick: Copilot packed those results into invisible Unicode characters inside a normal-looking link. You wouldn’t spot the data by reading it. If Copilot searches unrelated mail while handling an email, stop and report that email to security.

MITRE ATT&CK techniques

T1566.002 - Phishing: Spearphishing Link · PhishingT1656 - Impersonation

Similar attacks

Amatera Uses Fake AI Tools to Steal Data

Amatera Uses Fake AI Tools to Steal Data

Amatera is a subscription-based information stealer that affiliates distribute through fake AI-tool download pages, compromised websites, paid search ads, phishing, and deceptive ClickFix prompts. The scams convince people to copy and run a command to supposedly fix a CAPTCHA, browser error, or…

October 30, 2026
Phishers Exploit Microsoft 365 Outage

Phishers Exploit Microsoft 365 Outage

Attackers quickly used a real Microsoft 365 outage as cover for fake service-health emails. The messages warned that email held during the outage would be deleted within 24 hours unless the recipient clicked a link to restore it, likely aiming to steal credentials.

November 14, 2026
Salary Reviews and POs Fuel Phishing

Salary Reviews and POs Fuel Phishing

Microsoft observed real Q3 2026 phishing campaigns that used familiar workplace requests to steal credentials or deliver malware. One campaign targeted legal organizations with personalized salary-review documents containing QR codes, while another used purchase-order and payroll lures in disguised…

November 5, 2026
Five Active Phishing Lures Hit US Firms

Five Active Phishing Lures Hit US Firms

The article describes active campaigns that use familiar business documents, trusted file-sharing services, AI brands, and Microsoft sign-in flows to trick employees into granting account access. The most actionable patterns are phishing emails and document links that lead victims to Microsoft…

November 5, 2026
Campaign Spoofing Puts Donors at Risk

Campaign Spoofing Puts Donors at Risk

The FBI warns that scammers are impersonating political candidates and campaigns to solicit donations and steal personal information ahead of the midterms. Attackers use candidates’ names, logos, and AI-generated voices or messages, while weak email authentication can allow fraudulent messages to…

October 29, 2026
Stolen Microsoft 365 Tokens Hit 79 Organizations

Stolen Microsoft 365 Tokens Hit 79 Organizations

Attackers used fake Microsoft sign-in pages and device-code phishing to capture or obtain authenticated Microsoft 365 sessions, bypassing the need to repeat multi-factor authentication. They then took over accounts, added their own devices and authentication methods, and hid messages with inbox…

October 27, 2026