Design test
The lethal trifecta: a test for any AI tool
Data theft through an LLM tool needs three things together. Take any one away and the attack loses its route.
Last reviewed against NCSC, OWASP GenAI Security Project and Microsoft Security Response Center. Sources
The idea
Three capabilities
Simon Willison, writing on 16 June 2025, named the combination: a tool that can see your private data, that reads content from untrusted sources, and that can communicate with the outside world [6]. When one tool has all three, an attacker can plant instructions in the content, have the tool gather the private data and send it out [6].
The three
What each looks like at work
Private data
Mailboxes, shared drives, customer records, internal wikis or anything a retrieval feature can search.
Untrusted content
Inbound email, web pages, uploaded files and documents written by suppliers or the public. OWASP's indirect injection examples all start here [2].
Outward communication
Sending mail, calling APIs, loading images or links. In EchoLeak, data left through Teams and SharePoint URLs [4].
Applying it
Score each tool
Work through this for every assistant, agent and plug-in:
-
1
Mark yes or no for each of the three capabilities. -
2
Where all three are yes, treat the tool as high risk. -
3
Remove one capability: cut the data scope, filter or quarantine the content, or block the outward route. -
4
Record the decision and who accepted it.
Why not filter
The case against detection alone
Willison's advice is to avoid the combination entirely, and the post says nobody yet knows how to prevent these attacks with 100% reliability [6]. The NCSC makes a related point about vendors: be wary of any that claim to stop prompt injection [1].
Design beats detection
The NCSC favours deterministic, non-LLM safeguards that constrain what the system can do, rather than depending only on filtering content [1].
Next
Related reading
The prevention guide lists the controls that break each leg, and the plan puts them in order.
Keep reading
Read next
Vs SQL injection
Why the NCSC says prompt injection is not SQL injection, why parameterised queries have no equivalent, and what 'confusable deputy' means in practice.
Read the guide
Prevention
Prompt injection controls for UK teams: least privilege, human approval, separating untrusted content and logging, from OWASP, Microsoft and the NCSC.
Read the guide
EchoLeak case
EchoLeak, CVE-2025-32711, was a zero-click prompt injection flaw in Microsoft 365 Copilot. What happened, what Microsoft did and what it teaches.
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.