Prompt injection
Someone hides instructions in an email, document or web page, and your AI follows them instead of your staff.
AI security
Know what your AI can see, and what it can do.
AIFor.Wales runs a fixed-scope AI security review for SMEs and charities that already use chatbots, copilots or AI agents. We map where AI touches your data, threat-model it against the OWASP Top 10 for LLM applications, and give you a short written report that says what to fix first.
Your team's AI tools read your files, draft your emails and, increasingly, act on your behalf. We work out where that can go wrong and what to do about it, in plain English.
01Map it
Which AI tools you use, what data goes in, and who or what can act on the output.
02Threat-model it
Where each tool could be tricked, leak data or do more than it should.
03Fix it
Specific controls and supplier questions, in a report your board can read.
Why us
We build and run AI systems ourselves, so we know where they break.
AIFor.Wales is a software company. We run live AI systems every day, so the review starts from how AI behaves in practice: what a tool can really reach, what it does with a badly worded instruction, and what happens when a supplier changes something. It does not start from a checklist.
This is a review, not a penetration test. We read how your tools are set up, trace what they can reach, and assess the risks. We do not attack your systems.
- Agree the scope.The AI tools you actually use: chatbots, copilots, agents and anything built into software you already have.
- Map the data.What goes in, where it is stored, who can see it, and who or what can act on what comes out.
- Threat-model it.Assess each tool against the OWASP Top 10 for LLM applications, starting with the risks that matter most to a small organisation.
- Report and agree actions.A short written report with the controls to put in place, in priority order.
What we review
The risks that matter most to a small organisation.
We start from the OWASP Top 10 for LLM applications, a widely used reference for how AI-powered tools fail. These five are where we begin for SMEs and charities.
Sensitive data leakage
Personal, financial or case data ends up in a prompt, a log or an answer that the person asking should never see.
Excessive agency
An AI agent can send, delete, book or pay, with more permissions than the job needs.
Insecure output handling
What the AI writes is trusted and acted on without checking, so a wrong or manipulated answer becomes a real action.
Supply chain
The models, plug-ins and suppliers behind your tools, and what they can see and change.
Practical controls
Fixes you can put in place, not a list of worries.
For each risk we recommend the simplest control that works for an organisation of your size, and say who needs to own it.
What we recommend
Least-privilege permissions
Each tool and agent can reach only the files and actions its job needs.
Human review before actions
A named person checks anything that sends, changes, deletes or pays.
Logging
Enough of a record to see what the AI was asked, what it used and what it did.
Supplier due diligence
The questions to put to your AI suppliers about data, training, hosting, access and incident handling.
The report
Short, written, and useful to the people who decide.
You get a written report in plain English: the AI map, the risks found, the controls to put in place in priority order, and the questions to ask your suppliers. A trustee, funder or procurement team can read it without a glossary.
Next step
It feeds ISO/IEC 42001 if you want to go further.
The report can feed the risk assessment that the AI management standard ISO/IEC 42001 asks for, so the work is not repeated. We don't issue certificates; that is done by an accredited body. Read about our ISO/IEC 42001 readiness support.
Start here
Tell us which AI tools your team uses.
A short call, by phone or Teams, is enough to agree whether a review fits and what it would cover.
AIFor.Wales is a trading name of Blue Door Software Ltd, based in North Wales. Or email hello@aifor.wales. See our privacy notice.