AI and ML in Cybersecurity: UK Statistics and Threat Trends

Artificial intelligence and machine learning have moved from research labs into everyday security tooling faster than most IT teams have been able to plan for. AI and ML in cybersecurity now sit inside email filters, endpoint protection, fraud detection, and the systems that monitor network traffic for anything unusual, often without staff realising how much of their existing stack already depends on them.

The same technology is being picked up by attackers. The National Cyber Security Centre (NCSC) has said plainly that AI will almost certainly continue to make cyber intrusion operations more effective, and that threat actors are already using it to speed up reconnaissance, exploit development, and social engineering. That double use, by defenders and attackers alike, is why AI and ML in cybersecurity has become one of the more consequential shifts in the sector rather than another passing buzz term.

This article sets out what AI and ML actually mean in a security context, how UK organisations are using them for defence, how attackers are using the same tools against them, what UK regulators expect, and what to ask before buying any tool marketed as “AI-powered”.

What AI, ML and Deep Learning Mean in a Security Context

The three terms get used interchangeably in vendor marketing, which makes it harder to judge what a security product actually does. Artificial intelligence is the broad field of building systems that perform tasks normally associated with human judgment. Machine learning is a subset of AI in which a system improves at a task by learning patterns from data rather than following rules written by a person. Deep learning is a subset of machine learning that uses layered neural networks and often requires much larger datasets, which is why it tends to appear more in large vendor platforms than in smaller tools.

Term	What it does	Typical security use
Artificial intelligence	Broad term for systems performing tasks that normally need human judgement	Chatbots for security awareness, automated policy recommendations
Machine learning	Learns patterns from historical data to make predictions on new data	Spam and phishing filters, anomaly detection in network traffic
Deep learning	Neural network based machine learning, needs large datasets	Malware classification, image based threat detection, advanced behavioural analytics

For most small and medium UK organisations, the practical question is not which of these three categories a tool sits in. It is whether the vendor can explain, in plain language, what data the model was trained on and what happens when it gets something wrong. Understanding this distinction is the first real step towards properly evaluating AI and ML in cybersecurity, and the checklist later in this article sets out a short list of questions to ask before signing.

How AI and ML in Cybersecurity Are Used for Defence

Most defensive use of AI and ML in cybersecurity falls into three overlapping roles: catching threats faster than a human analyst could on their own, determining which of many alerts actually deserve attention, and reducing the volume of phishing and social engineering that reaches staff inboxes.

Automated Threat Detection and Response

Machine learning models are trained on historical logs of normal network behaviour, which allow them to flag activity that falls outside that baseline, such as a sudden spike in outbound data from a single device or a login attempt from an unfamiliar location. This kind of automated flagging is one of the clearest, most established applications of AI and ML in cybersecurity, and it is the same principle behind endpoint detection and response tools, which increasingly bundle machine learning into their alerting so that suspicious processes are flagged and, in some cases, isolated automatically before a human ever reviews the log.

The main benefit of applying AI and ML in cybersecurity this way is speed. A model can review far more log data per minute than any analyst, and it does not get fatigued during a long shift. The trade-off is that a model trained on the wrong baseline, or trained on too little data, will either miss real threats or bury a security team in false alarms.

Predictive Analytics and Risk Prioritisation

Predictive analytics uses AI and ML in cybersecurity to analyse patterns across past incidents and flag which systems or user accounts pose the highest risk right now, rather than treating every asset as equally exposed. This matters because most security teams, particularly in small and medium UK businesses, do not have the staff to investigate every alert with the same depth.

A well-tuned model can help a two-person IT team behave more like a much larger one, by pointing them towards the small number of alerts that carry real risk instead of the much larger number that do not. This is where AI and ML in cybersecurity earn their keep for smaller teams, but it cannot replace the judgment calls that follow, such as deciding whether a flagged account belongs to someone on annual leave or someone whose credentials have actually been compromised.

Phishing and Social Engineering Defence

Phishing remains the most common way UK organisations are attacked. DSIT’s Cyber Security Breaches Survey 2025/2026 found that 85% of businesses and 86% of charities that had a breach or attack in the past year had faced at least one phishing attempt, and the accompanying interviews recorded growing concern among organisations about AI-assisted impersonation, making these messages harder to spot.

Machine learning-based email filters look beyond obvious spam signals to catch subtler indicators, such as a sender address that mimics a real supplier or unusual invoice wording, and this is one of the more mature and widely deployed uses of machine learning in cybersecurity today. Even so, no filter catches everything, which is why staff training remains part of any sensible approach to AI and ML in cybersecurity rather than a fallback for when the technology fails.

The Other Side: How Attackers Use AI

The same capabilities that make AI and ML in cybersecurity useful to defenders also make them equally useful to attackers. In its report on the impact of AI on the cyber threat to 2027, the NCSC states that this will highly likely increase the volume and impact of intrusions by making existing attack methods faster and cheaper to run, rather than by inventing entirely new ones. That distinction matters for how UK organisations plan their defences: the threats themselves are largely familiar, but they are arriving more often and from a wider range of attackers who previously lacked the skill or time to run them at scale.

Generative AI and Automated Phishing

Attackers using AI and ML in cybersecurity offensively have found generative AI particularly useful. These models can produce fluent, well-targeted phishing messages at a scale that would previously have needed a team of writers, and can tailor wording to a specific company or individual using information scraped from public sources such as LinkedIn or a corporate website. The NCSC has confirmed that threat actors are already using AI to enhance reconnaissance, basic malware generation and the exploitation of victims through social engineering.

For UK organisations, the practical effect is that the traditional advice to look for spelling mistakes or awkward phrasing in a phishing email is far less reliable than it used to be. Detection increasingly depends on checking sender authenticity and unusual request patterns rather than surface-level writing quality, which is one reason why guidance on how to avoid malware attacks has shifted its emphasis in recent years.

Data Poisoning and Attacks on Model Integrity

A newer risk targets the AI systems themselves rather than the humans using them. Data poisoning involves feeding a machine learning model corrupted or misleading training data so that it learns the wrong patterns, for example so it fails to flag a particular type of malicious traffic as suspicious.

The NCSC’s Guidelines for secure AI system development flag data poisoning and prompt injection as novel vulnerabilities that need to sit alongside standard cybersecurity threats when an organisation is assessing risk, rather than being treated as a separate, lower-priority concern.

This matters most for organisations building or fine-tuning their own models, but it is relevant to anyone buying a third-party AI security tool too, since the vendor’s own model is only as trustworthy as the data and processes behind it. Any organisation assessing AI and ML in cybersecurity risk needs to weigh this alongside the more familiar threats covered elsewhere in this article.

UK Regulation and Compliance for AI in Cybersecurity

There is no single UK AI law, including rules written specifically for AI and ML in cybersecurity. Unlike the EU AI Act, the UK has stuck with a principles-based approach set out in the 2023 white paper and developed since through guidance from existing regulators rather than new primary legislation, and organisations that expect a single rulebook are often surprised to learn there isn’t one.

In practice, three sources matter most for a UK organisation using AI in a security context. First, the NCSC and the Department for Science, Innovation and Technology (DSIT) jointly published an AI Cyber Security Code of Practice, which served as the basis for a new global technical standard issued by the European Telecommunications Standards Institute (ETSI) and endorsed internationally.

Second, the Information Commissioner’s Office (ICO) continues to develop guidance on automated decision-making, particularly following the changes introduced by the Data (Use and Access) Act 2025, which affects any AI security tool that makes decisions about individuals, such as automatically locking an account or flagging a person as a risk. Third, existing UK GDPR obligations apply in full to any AI system that processes personal data as part of threat detection, meaning organisations still need a lawful basis, clear retention limits and a way for someone to challenge an automated decision that affects them.

The GDPR implications of AI-powered monitoring tools are often underestimated. A system that scans staff email metadata for anomalies, for instance, is processing personal data regardless of how the vendor markets the tool, and that processing needs the same lawful basis and documentation as any other monitoring activity. Getting this right is as much a part of deploying AI and ML in cybersecurity responsibly as choosing the right model in the first place.

Will AI Replace Cybersecurity Professionals?

This is the question UK security teams ask most often, and the honest answer is less simple than a flat yes or no. The NCSC’s own guidance on using AI for vulnerability discovery states directly that AI models enhance the capabilities of cybersecurity staff rather than replace them, and notes that simply finding more vulnerabilities does not improve security if there is no process in place to triage and fix what the tool surfaces.

What is shifting is the type of work analysts spend their time on. Manual log review and first line alert triage are increasingly handled by machine learning models, which frees analysts to spend more time on investigation, judgment calls and the parts of incident response that still need a person, such as deciding how to communicate a breach to customers or regulators. The cybersecurity skills gap in the UK remains real, and the rise of AI and ML in cybersecurity is, if anything, making that gap more visible rather than less, since organisations still need staff who understand what the models are doing well enough to spot when they get it wrong.

Questions to Ask Before Buying AI Security Tools

Vendor marketing for AI-powered security products rarely explains what is actually happening under the surface, and a short evaluation checklist helps separate genuine capability from a rebadged rules engine. Before signing a contract for any product marketed around AI and ML in cybersecurity, it is worth asking the vendor to answer each of the following in writing rather than verbally in a sales call:

• What data was the model trained on, and does that include any UK-specific threat data, or is it trained mainly on US-sourced datasets?
• How does the system behave when it encounters something outside its training data, and what happens to the false positives it generates?
• Can a human reviewer see the reasoning behind a flagged alert, or does the system only return a risk score with no explanation?
• What happens to the data the tool collects from our systems? Where is it processed and stored, and for how long is it retained?
• Does the vendor have a documented process for updating the model as new threats emerge, and how often does that happen?
• How does the tool fit with our existing obligations under UK GDPR, including the right to a human review of automated decisions?

Any vendor unable to answer these clearly is not necessarily using AI badly, but the inability to explain the basics is itself a risk worth weighing against whatever the tool promises to catch.

Challenges of Implementing AI in Cybersecurity

Cost is the most immediate barrier for many UK organisations. Licensing an AI-powered platform, training staff to interpret its output, and maintaining the integration with existing tools all add up, and the ongoing cost of keeping a model current against new threats does not stop once the initial deployment is finished. For a small or medium-sized organisation already stretched on data breach prevention budget, an AI security platform is rarely the first purchase that makes sense, and it works best layered on top of the basic controls covered in that guide rather than in place of them.

Bias and data quality are a closer, more technical problem for anyone deploying AI and ML in cybersecurity. A model trained on a narrow or unrepresentative dataset will produce unreliable results, whether that means missing genuine threats that do not match its training patterns or repeatedly flagging normal behaviour from a particular team or system as suspicious. Insufficient or poor quality training data compounds this, since a model can only be as accurate as the data used to build it, and correcting a poorly trained model after deployment is far harder than getting the training data right from the outset.

Overreliance on automated tooling is a related risk. Treating an AI system’s output as a final answer rather than as one input into a human decision removes exactly the kind of judgment that catches edge cases a model has never seen before. The safest position most UK security teams land on is to treat AI and ML in cybersecurity as a way to extend what analysts can cover, not as a replacement for the analysts making the final call, which echoes the NCSC’s own framing of AI as accelerating skills rather than substituting for them.

Finally, the ongoing cybersecurity skills gap means many UK organisations, particularly smaller ones, do not have anyone in house who fully understands how their AI security tools work. That makes it harder to spot when a tool is underperforming, harder to challenge a vendor’s claims with any confidence, and harder to build the wider cybersecurity principles that AI tooling is supposed to support rather than replace.

If your organisation is evaluating AI and ML in cybersecurity tools, work through the checklist in this article with the vendor directly, and treat a lack of clear answers as useful information in its own right.

FAQs