AI Security Risks for Small Businesses: The Five That Actually Bite

Most writing about AI risk is about the wrong things. Rogue systems, mass unemployment, machines that deceive their operators. Interesting questions, and not what is going to cost your business money this year.

Here are the five that actually happen to UK small businesses, in order of frequency.

This article is a companion to the risks of AI for small businesses.

1. Staff pasting confidential information into public tools

The most common AI incident in British business, by a wide margin, and it is not close.

Someone has a deadline. They paste a client document, a draft contract, a spreadsheet of figures or an HR matter into a consumer AI tool, because it is faster. Depending on the tool and the tier, that content may be retained and may be used for training.

Why it keeps happening: most staff genuinely do not know that the free and business tiers of the same product behave differently. They assume the tool works like a search engine.

What actually fixes it: approve a business-tier tool and provide it, publish a concrete list of what may never be pasted anywhere, and explain the tier difference once, properly. A ban does not fix it, it only stops you hearing about it. See shadow AI.

2. Much better phishing

AI has removed the two things that used to give phishing away: poor English and generic content.

A current attack reads like a real message from a real supplier, references a real project, and arrives at a plausible moment. Voice cloning has made the "urgent call from the finance director" attack considerably more effective too.

What actually fixes it: process, not vigilance. A rule that payment details are never changed on the basis of an email or a call alone, verified by a call back on a previously known number. That single control defeats nearly all of it, and it is free.

Telling staff to look out for suspicious emails no longer works, because they are no longer suspicious.

3. Your chatbot saying something you have to honour

If a customer-facing bot states a price, a term, a delivery time or a policy, your business made that statement. Consumer protection law does not care that a model generated it.

What actually fixes it: restrict the bot to published information, make it escalate on anything about price, terms or complaints, and review transcripts weekly for the first few months. Weekly transcript review is the control almost nobody does and the one that finds the problem before a customer does.

4. Data going somewhere you did not agree to

Distinct from the first risk, because it is about tools rather than people. Software you already use adds AI features in an update. Those features may process your data differently from the rest of the product, and the terms may have changed.

What actually fixes it: an annual review of terms on the handful of tools holding your important data, and a check of what AI features have been switched on by default. Most businesses have never done either.

5. Over-trusting output in a decision that matters

Not a security risk in the traditional sense and it belongs here because the consequences are comparable.

A financial model, a legal position, a technical specification or a compliance answer, produced quickly, checked lightly, acted upon. The output was fluent, the deadline was real, and nobody with subject knowledge looked at it properly.

What actually fixes it: a rule tied to consequence rather than to volume. Anything that leaves the building, goes to a client, or informs a decision gets checked by someone who knows the subject. See AI literacy for non-technical leaders.

What we would not spend money on

AI security platforms, for a business under a few hundred people. The five risks above are addressed by decisions, rules and one training session, not by a product.

Network monitoring to detect AI use. It finds less than asking does, and it damages the conversation you need to have.

Insurance specifically for AI risk, before checking whether your existing professional indemnity and cyber cover already responds. Usually it does, and the sensible step is to ask your broker rather than buy something new.

Key Takeaways

• Staff pasting confidential information into public tools is the most common real incident by a wide margin, and it is a knowledge gap rather than a discipline problem.
• AI has made phishing genuinely convincing. Vigilance no longer works; a call-back rule on payment changes does, and it is free.
• Weekly review of chatbot transcripts is the control almost nobody runs and the one that catches problems before customers do.
• Check what AI features your existing software switched on by default, and review terms annually on the tools holding important data.
• For a business under a few hundred people, none of these need a product. They need decisions, two rules and one training session.

Frequently Asked Questions

Do we need an AI security policy?

You need an AI use policy, which covers this and more, and it should be two pages rather than fifteen. Template in writing an AI use policy for a small business.

Is it safe to use AI tools at all?

Yes, with the business tier and a clear rule about what goes in. The greater risk for most businesses is unmanaged use on consumer accounts, which is what happens when the official answer is no.

What about AI writing our code?

Two real risks and neither is exotic: credentials being pasted into a prompt, and code reaching production without review. Both are addressed by rules you probably already have for human-written code. Apply the same ones.

How would we even know if something had gone wrong?

Mostly you would not, which is the argument for making near-miss reporting safe and normal. A business where someone will say "I think I pasted something I should not have" within the hour is in a far better position than one that has never had a report.

---

Want a plain look at where your actual AI exposure sits? Talk to Halo Technology Lab. Our strategy and scoping service covers the five above, and we will tell you when the answer is that you do not need to buy anything.