November 4, 2026Mark Hayward

Today's BLOG: Cyber Security AI and ISO 42001 Standard v2 (Series 1a) - Core AI Techniques: Machine Learning, Deep Learning and NLP

How machine learning, deep learning and natural language processing each contribute to detection, triage and response, and how to combine them without over-trusting them.

Mark Hayward

Mark Hayward

Cyber Security Expert · UK Armed Forces Veteran · 24+ years experience

Three techniques do most of the work when AI is applied to security: machine learning, deep learning and natural language processing. They are often described as if they were interchangeable, but each solves a different kind of problem, and knowing which is which prevents a great deal of wasted effort.

Machine learning is the workhorse. A model learns a relationship from examples, then applies it to new data. In security that shows up in anomaly detection, where a baseline of normal behaviour is established and deviations are scored, in malware classification, where features of a binary predict its family, and in user behaviour analytics, where a compromised account looks different from its owner. The appeal is that the model generalises beyond rules that were written before the attack existed. The cost is that it needs representative data and ongoing retraining, because a model trained on last year's traffic quietly stops being right.

Deep learning extends machine learning with neural networks that have many layers. It is strongest where the input is unstructured: raw network packets, executables, documents, images, or long stretches of text. This makes it valuable for detecting attacker tooling that has been recompiled to defeat signature matching, for analysing command lines at scale, and for classifying documents during incident response. Deep models also demand more compute, more data and more scrutiny, because their internal reasoning is harder to explain to an auditor or a court.

Natural language processing handles the text that security work generates in enormous volume. Incident notes, threat intelligence reports, phishing emails, vulnerability advisories and chat logs are all language data. NLP can extract indicators of compromise, summarise a fifty page vendor report into the five facts that matter, group related tickets so an analyst sees one campaign rather than two hundred alerts, and detect social engineering patterns in inbound email. For threat intelligence specifically, it converts a pile of feeds into a trend line a human can act on.

Putting the three together is where real capability emerges. A pipeline might use NLP to parse a threat report, machine learning to score the relevance of each indicator to the organisation's estate, and deep learning to check whether matching binaries appear in the environment. Detection, enrichment and prioritisation become one flow instead of three disconnected tools, and response time falls accordingly.

Techniques also support action, not only analysis. Once an incident is confirmed, models can rank affected assets by business criticality, propose containment steps, or simulate the likely consequences of different responses. That is where AI changes the economics of defence: the expensive part of an incident is usually the delay between detection and decisive action.

The limits deserve equal billing. Models drift as the environment changes, they can be manipulated through adversarial input or poisoned training data, and they reproduce whatever bias exists in their training sets. Explainability is not a luxury when an alert triggers an outage or a personnel decision. Sensible practice treats model output as corroborated evidence, keeps a human accountable for consequential actions, and documents how each automated judgement was reached.

For teams starting out, the practical advice is to pick one technique for one problem and measure it honestly. A well validated anomaly detector that reduces alert volume by a third is worth more than an ambitious platform nobody trusts. Build from there, keep the data pipeline clean, and revisit the model whenever the threat landscape or the network changes.

📚 Want to go deeper?

Cyber Security AI and ISO 42001 Standard v2 (Series 1a)

The book behind this series — how AI is applied to cyber defence, and how the ISO 42001 management system keeps it governed, evidenced and auditable. Now in Hardcover, Paperback and eBook.

🎁

Get Chapter 1 free

Subscribe and download the opening chapter of Cyber Security for Beginners — plus new book alerts and expert insights.