Cyber Security for Businesses That Aren't Banks

Most security advice is written for organisations with a security team. If you have one person who "also does IT", that advice is not just unhelpful — it is actively harmful, because it buries the five things that would genuinely protect you under two hundred that would not.

This guide is the opposite. It covers what actually gets exploited, in the order it gets exploited, and what stops it.

How attacks actually start#

Ignore the film version. Real breaches in ordinary businesses begin in a small number of ways:

Stolen or reused passwords. Someone used their work password on a site that got breached. Attackers try the combination everywhere. No hacking required — they simply log in.

Phishing. An email that looks like your bank, your supplier or your own IT desk. The modern version is not badly spelled; it references a real invoice number and arrives during a real project.

Unpatched software. A known flaw in software you run, with an exploit published weeks ago. Scans find your server automatically. Nobody targeted you specifically.

A supplier. Your accounting platform, your web host, a contractor's laptop. You inherit their security whether you assessed it or not.

A departing employee's access. The account nobody disabled.

Notice what is absent: sophisticated zero-days, nation-state tooling, anything requiring genius. Ordinary businesses are compromised by ordinary means, opportunistically.

That is good news, because ordinary means have well-understood defences.

The controls that do the most work#

If you do nothing else, do these. They are ranked by protection per unit of effort.

1. Multi-factor authentication, everywhere it is offered. This single control defeats stolen passwords, which is the most common entry route. Email first, then anything financial, then everything else. Prefer an authenticator app or a passkey over SMS — SIM swapping is real, though SMS still beats nothing by a wide margin.

2. Patch quickly, and know what you run. You cannot patch what you have forgotten. Keep a list of your systems, who owns each, and where updates come from. Enable automatic updates where you can tolerate them. The window between a flaw being published and being mass-exploited is now days.

3. Back up, and test restoring. Backups are worthless until proven. The standard is three copies, two kinds of storage, one off-site and offline or immutable — ransomware deliberately encrypts backups it can reach. Restore something real, on a schedule. Most failed restores are discovered during the incident.

4. Least privilege. People and systems get the minimum access to do their job. Administrator rights for everyone means one phished account compromises everything. This is free and it is mostly a discipline problem.

5. Off-board on the day. A leaver's access ends the day they leave, not the month after.

6. Verify payment changes out of band. A supplier emailing new bank details is the most reliably profitable attack there is. Phone the number you already had on file — never the number in the email. This one control prevents a category of loss that insurance often disputes.

Six controls. Together they address the large majority of what actually happens to organisations your size. Everything below is refinement.

Where the money goes wrong#

Two failure patterns, both expensive.

Buying tools instead of fixing basics. A threat-detection platform on an estate without MFA is a very good way of watching an attacker log in with valid credentials. Tools are a multiplier on fundamentals, not a substitute.

Compliance as a proxy for security. Passing an audit means you documented controls on the day of the audit. Useful for winning contracts; not the same as being hard to break into. Treat the certificate as a business requirement and security as a separate, ongoing one.

What AI changes#

Genuinely new ground, and worth separating what has actually changed from what merely sounds new.

Phishing got better, cheaply. The grammatical errors that used to give phishing away are gone. Attackers can now produce fluent, contextual, personalised messages at volume — including voice. Defence: stop training people to spot bad English, and move to process — verify requests for money or credentials through a second channel, always, regardless of how convincing the message is.

Your staff are pasting company data into AI tools. Not maliciously — to get work done. That data may be retained or used for training depending on the service and plan. Defence: provide an approved tool with appropriate terms. A blanket ban does not stop it, it just makes it invisible.

AI systems create a new attack surface. If you have built anything with a language model that reads external content and can also take actions, you have a prompt injection exposure: hostile instructions hidden in a document or web page that the model follows. This is not solved by better prompts. It is contained by giving the system narrow tools, requiring human approval for anything outbound or destructive, and never granting one system both broad read access and broad write access. Our AI Agents guide covers the control design in detail.

Deepfake authorisation. A video call or voice message from a "director" approving an urgent payment. Defence: the same out-of-band verification as above. Authority must be confirmed through a channel the attacker does not control.

What has not changed: attackers still mostly log in with valid credentials, and MFA still stops that.

Where to look for the detail. Two published lists cover this ground properly and are worth handing to whoever builds your AI systems: the OWASP Top 10 for LLM Applications (2025) — prompt injection, sensitive information disclosure, supply chain, system prompt leakage, excessive agency — and the OWASP Top 10 for Agentic Applications (2026), published in December 2025 for systems that take actions rather than answer questions. If you are commissioning AI work, asking a supplier which of these they design against is a fast way to find out whether they have thought about it at all.

On obligations: EU AI Act transparency requirements applied from 2 August 2026 — people must be told when they are interacting with an AI system, and generated content must be labelled. The heavier high-risk obligations were deferred to December 2027. See our AI guide for what that means in practice.

When something goes wrong#

Decide these things now, because you will not think clearly during an incident.

• Who decides? One named person with authority to disconnect systems and spend money.
• How do you communicate if email and chat are compromised? Agree a fallback channel.
• Who do you call? Your insurer, your legal contact, your IT provider — numbers stored somewhere that does not depend on the systems being up.
• What are your obligations? Many jurisdictions require notifying a regulator within a short, fixed window. Know yours before the clock starts.
• Preserve evidence. The instinct is to wipe and rebuild. Snapshot first — you may need to prove what did and did not leave.

Write this on one page. A one-page plan people have read beats a fifty-page plan nobody has opened.

How much should you spend?#

A reasonable way to decide, without a risk framework:

Estimate what a bad week costs — systems down, data disclosed, payments diverted, customers notified. Then ask what proportion of that you would pay annually to make it substantially less likely. That number is your budget, and for most organisations the first slice of it buys MFA, backups, patching and training. Those are cheap. The expensive items should be justified against what they add on top of a solid base.

Be honest about your actual exposure. A ten-person consultancy holding client documents faces different risks from a payments business. Copying a bank's control set will bankrupt you and still miss the thing that gets you.

FAQ#

We're too small to be a target. Is that true?#

No — and it is the most expensive misconception in this area. Most attacks are not targeted. They are automated scans looking for any system with a known flaw or a reused password. Being small does not make you invisible; it usually just means fewer defences.

Is multi-factor authentication really that important?#

Yes. If you implement one thing from this guide, implement MFA on email first. Stolen credentials are the most common way in, and MFA is the control that breaks that chain. Email specifically, because email resets every other password you own.

Is antivirus enough?#

No, but it is not useless either. Modern endpoint protection catches commodity malware. It does not help when an attacker logs in with a valid password, which is the more common case. Treat it as one layer.

Should we pay a ransom?#

Plan so you never have to decide under pressure — tested, offline backups are the answer. Payment funds the industry, does not guarantee recovery, and in some jurisdictions may be unlawful depending on who receives it. Take legal advice before anything else.

Do we need a security certification?#

If customers or regulators require it, yes — it is a commercial requirement. Do not confuse it with being secure. Achieve the controls first; the certificate then documents something real rather than describing something aspirational.

Is it safe to use AI tools at work?#

With reasonable care, yes. Use business-tier services whose terms exclude training on your data, decide what categories of information are off-limits, and tell people clearly. The realistic alternative to an approved tool is not "no AI" — it is staff using consumer accounts you cannot see.

How often should we test our backups?#

Quarterly at minimum, and after any significant change. Test a real restore of something that matters, not just that the backup job reported success. The gap between "backups are running" and "we can recover" is where organisations discover they had neither.

What else is coming for Cyber Security