How Can Generative AI Be Used in Cybersecurity? A UK SME Guide

How can generative AI be used in cybersecurity? On the defensive side, it helps analysts and IT teams read, write and summarise faster. It can condense a long incident into a plain-English summary, turn a question into a database query, explain a suspicious script and draft a report or policy. On the offensive side, criminals use the same technology to write more convincing phishing messages and to imitate voices and faces.

For a UK SME the opportunity is real but narrow. These tools reduce the time spent on reading and writing around security work. They do not replace the judgement needed to decide what an alert means or what to do about it, and they introduce risks of their own that need managing.

This guide sets out the main defensive uses, the misuse to expect, the specific risks, and a short governance checklist. If you want security support that already includes sensible tooling, our managed IT support team can advise on what fits your size.

What generative AI adds to security work

Generative AI tools produce text, code or other content from a prompt, using models trained on very large data sets. Security teams spend a great deal of time on tasks that are mostly language: reading logs and alerts, writing queries, explaining findings to non-specialists and producing documents. That is where these tools fit.

They are weaker at tasks that need certainty. A model can sound confident and be wrong, so it suits first drafts and analysis support, not final decisions. Keep that distinction in mind through the rest of this article.

How can generative AI be used in cybersecurity defensively?

SOC copilots

Microsoft Security Copilot is a generative AI security product that accepts natural-language prompts, with a standalone experience and embedded experiences in Microsoft security products including Defender XDR, Sentinel, Intune and Entra, plus integrations with some third-party services. Microsoft's documentation lists uses such as summarising complex alerts into actionable context, translating plain language into KQL queries, analysing suspicious scripts, drafting and checking security policies and producing audience-specific reports. As at October 2026, Microsoft notes it is not designed for US government clouds, and licensing details should be confirmed with Microsoft or a partner. See Microsoft Learn on Security Copilot.

Alert summarisation

An analyst facing an incident with dozens of related events can ask for a timeline and a summary, then verify it against the source evidence. This saves reading time, particularly out of hours.

Query generation

Writing queries in a specialist language slows down many IT generalists. A tool that drafts the query lets them search for sign-in anomalies or file activity without memorising syntax. Run the query, read the results and check the logic rather than trusting the first attempt.

Phishing analysis

Models can explain why an email looks suspicious: mismatched sender, urgent tone, unusual link structure. That helps staff learn too. Our guide to phishing types and prevention covers the underlying techniques.

Report and policy writing

First drafts of incident reports, board updates, acceptable use policies and supplier questionnaires are a sensible use. Always have a person edit them for accuracy and for your actual practice, because a policy that describes controls you do not operate creates its own risk.

Three everyday examples for a small IT team

The examples below show where the time saving comes from and where the checking step sits.

• A suspicious sign-in. The analyst asks for a summary of the related events and a suggested query to find other sign-ins from the same address. The tool saves reading time; the analyst confirms the evidence and decides whether to reset the account.
• A reported email. A member of staff forwards a message that looks wrong. The tool lists the features that raise suspicion. IT still checks headers and links with proper tooling before blocking anything.
• A supplier questionnaire. A customer asks 80 security questions. The tool drafts answers from your existing policies. A named person corrects anything that does not match what you actually do, because inaccurate answers can cause contractual problems.

In each case the tool shortens the preparation and a person owns the outcome. If a use case would have no human checking the result, treat that as a warning sign and redesign it.

Offensive misuse: what attackers do with the same tools

Generative AI lowers the effort needed to write fluent messages. Poorly worded phishing used to be a useful warning sign, and that signal is now weaker. Attackers can produce clean, well-targeted emails in any language and adapt them to a recipient's role.

• AI-written phishing and fraud emails that read naturally and reference real business context.
• Deepfake voice and video used to impersonate senior staff and push through payments or access changes.
• Faster research and scripting by lower-skilled attackers.

The defence is mostly process, not detection. Agree call-back rules for payment and bank detail changes, and require a second approver for unusual requests. Executives are the usual targets, a tactic known as whaling. Staff awareness also needs to reflect that polished writing no longer proves an email is genuine; see our guide to cybersecurity awareness training.

The risks of using generative AI for security

Data leakage

Anything pasted into a public AI tool may leave your control. Logs, customer data, credentials and incident details do not belong in consumer chatbots. OWASP lists sensitive information disclosure as one of its top risks for large language model applications.

Hallucination and misinformation

Models can produce plausible but wrong answers, including invented commands or incorrect remediation steps. OWASP also lists misinformation among its top ten risks. Verify before acting.

Prompt injection

Prompt injection happens when untrusted content, such as text in an email or a web page, is treated by the model as instructions. The NCSC explains that large language models do not inherently separate instructions from data, that prompt injection will probably never be fully prevented, and that the goal is to reduce likelihood and impact. Its advice includes not giving a model that handles untrusted input access to privileged tools, using deterministic controls to constrain actions, being sceptical of products claiming to stop it outright, and logging inputs, outputs and tool calls. Read the NCSC blog on prompt injection.

Excessive agency

The OWASP Top 10 for LLM Applications 2025 also includes excessive agency, where a model is allowed to take actions beyond what is safe. In security work that means being cautious about letting an assistant change firewall rules, disable accounts or send messages without approval.

Governance checklist for SMEs

You do not need a heavy programme. A one-page standard that everyone follows is better than a long document nobody reads.

• Approved tools only. List which AI services staff may use for work and block or discourage the rest.
• Data rules. State what must never be entered: personal data, credentials, client files, unreleased financials and incident evidence.
• Check data access. Assistants that read your files and mail are only as safe as the permissions behind them. Fix over-shared files and permissions first. Tools such as Microsoft Purview help with classification and protection.
• Human review. Name who checks outputs before use, and who approves automated actions.
• Logging. Keep records of prompts and actions where the product allows, so you can investigate problems.
• Supplier checks. Ask vendors where data is processed, whether it is used to train models and how it is retained.
• Incident route. Decide what staff do if they paste something they should not have.
• Review date. Revisit the standard at least every six months because products change quickly.

How to start safely

1. Begin with low-risk drafting. Policy first drafts, report summaries and awareness materials carry little operational risk.
2. Tidy permissions before enabling assistants that read your mailboxes and files; a cyber security audit is a good way to find over-shared data. Our overview of Microsoft 365 Copilot explains how it works with your Microsoft 365 data.
3. Pilot with a small group for a few weeks and record what saved time and what was wrong.
4. Add investigation support only once the team can check outputs against source data.
5. Fit it into a wider plan. Our AI roadmap for business guide shows how to sequence adoption across the organisation.

Baseline security still does most of the protective work. Patching, multi-factor authentication, endpoint protection (see our managed EDR services) and tested backups matter more than any assistant, and our cybersecurity for small business page covers them.

Conclusion: generative AI in cybersecurity needs oversight

So, how can generative AI be used in cybersecurity? Defenders can use it to summarise alerts, write queries, explain scripts and draft documents, saving time on language-heavy work. Attackers use it to sharpen phishing and impersonation, so verification processes matter more than ever. The sensible approach is to start with low-risk tasks, protect your data, treat output as a draft and keep a person responsible for decisions. Product capabilities and guidance move quickly, so recheck vendor and NCSC material as at October 2026.

For help choosing and governing these tools, speak to NetMonkeys.

Frequently Asked Questions (FAQs)