Case study
EchoLeak: a prompt injection case study
An email that nobody opened could, via Copilot, pull out a user's data. Microsoft fixed EchoLeak before customers had to act, and the design lesson remains.
Last reviewed against NCSC, OWASP GenAI Security Project and Microsoft Security Response Center. Sources
The facts
What was reported
EchoLeak is tracked as CVE-2025-32711 and was rated critical, with a CVSS score of 9.3 [4]. Aim Security found it and reported it to Microsoft [4]. Microsoft's advisory described it as AI command injection in M365 Copilot that allows an unauthorised attacker to disclose information over a network [4].
-
1
An attacker sends an email with a malicious prompt hidden in markdown. -
2
Later, a user asks Copilot a routine business question. -
3
Copilot's retrieval engine mixes the hostile email with the user's sensitive data. -
4
The data leaves through Microsoft Teams and SharePoint URLs. -
5
No click or other user action is required.
Aim Security described it as a "zero-click AI vulnerability" [4].
Outcome
Fixed, with no known abuse
Microsoft said it had already addressed the flaw and that customers needed to take no action [4]. It was added to the June 2025 Patch Tuesday list [4]. The report found no evidence of malicious exploitation in the wild [4].
Why it still matters
A fix for one chain does not remove the class. The NCSC's position is that today's LLMs keep instructions and data in one undivided stream, so a complete mitigation may never exist [1].
Lessons
Read it against your own tools
Mapping to the trifecta
The case combined untrusted content (the email), private data (the user's files and mail) and an outward route (the URLs). That is the lethal trifecta [6].
Layered defences
Microsoft's later write-up describes limiting damage through data access controls, deterministic blocking of exfiltration techniques and approval for sensitive actions [3].
For the steps that follow, see the prevention guide.
Keep reading
Read next
What is it?
What prompt injection means, in plain English: how text fed to an LLM can override its instructions, and why UK organisations should care.
Read the guide
Can it be fixed?
Can prompt injection be fixed? The NCSC and OpenAI both say probably not completely, so the aim is to cut likelihood and impact. What that means for you.
Read the guide
Do filters work?
Do prompt injection filters, deny-lists and detection products work? What the NCSC, OWASP and Microsoft say about their value and their limits.
Read the guideHow CyPro can help
Need a hand planning for this?
Understanding the change is one job. Working out what it means for your systems, your suppliers and your roadmap is several more. CyPro's consultants can take any of them on.
Fractional CISO for the board
A senior security lead who advises the board on what to fix first, what to accept and what to tell customers, and owns the decision record.
More on cypro.co.uk
Hardening and segmentation
Our consultants put the controls in place that limit the damage when something is missed: firewall rules, network isolation and tighter admin access.
More on cypro.co.uk
Penetration testing
A focused test of your systems and the controls around them, so you know they hold before an attacker finds out for you.
More on cypro.co.uk
Vulnerability scanning
Find every internet-facing system you run, including the forgotten ones, and see which are still on a vulnerable build.
See the scanning service
24/7 security monitoring
Our SOC analysts watch your systems day and night for the signs of an attack in progress: unexpected processes, new outbound connections, logins that do not fit.
More on cypro.co.uk
Incident response
If you may already be compromised, our responders contain it, preserve the evidence, find out what was touched and get you running again safely.
More on cypro.co.uk
Plan ahead
Want a plan for your organisation?
Prompt injection only shows up when someone tries it against your own tools. Penetration testing can probe the AI features you run or build, to see what a hostile instruction could reach.