How Can Generative AI Be Used in Cybersecurity?

A security operations center, the team that watches a company’s systems around the clock, drowns in raw material. Alerts fire all day. Logs pile up by the million. Threat reports run to dozens of pages. The analysts on shift are skilled, but there are never enough of them, and most of a shift goes on reading and writing rather than judging. Generative AI, systems that create new text and code instead of only sorting old data, has found its first real security jobs in that gap.

This guide takes the question in order: the defensive work first, then the uncomfortable truth that attackers use the same tools, then the limits: confident wrong answers, sensitive data carried away, and attacks aimed at the AI itself. One rule runs through all of it. A person stays accountable for every decision.

Two kinds of AI, and one important difference

Security tools have used machine learning for years, long before chatbots arrived. Those older models classify and score: an email as spam or clean, a login as normal or odd, a file as known good or known bad. They answer with a judgment that is right or wrong about one signal at a time.

Generative AI does something else. It produces new material: a summary of an incident, an explanation of a script, a draft of a detection rule, a query written from a plain question. That output reads like a junior analyst’s work: fast, fluent, confident. It also fails like one. A classifier gives a wrong score. A generator can hand you a smooth paragraph with an invented fact buried inside it, stated as calmly as the true parts. Every use below works because a person checks the draft before it turns into action.

Alert triage and incident summaries

Triage is the first job of every shift: deciding which alerts deserve a human hour and which are noise. A busy center receives far more alerts than its analysts can read closely, so the sorting decides what gets seen. Generative AI now does the first pass. It gathers the alerts that belong to one incident, pulls in the context, and writes a short summary: what happened, which machines and accounts are involved, and what looks urgent. The analyst starts from a brief instead of a pile.

The pattern already ships in real products. Microsoft documentation for Security Copilot describes exactly this work: triaging complex alerts into actionable summaries, with step by step guidance responders can accept or reject. The UK National Cyber Security Centre makes the same point, noting that AI can improve the detection and triage of attacks and help spot malicious email. The value is reading speed, applied to work that used to eat the first hour of every investigation.

The summary also becomes the record. When an incident closes, the team needs a timeline and a write-up for managers, auditors or customers. The AI drafts both from the case notes, the analyst corrects them, and reports that once waited days now leave with the shift.

Threat intelligence, digested

Threat intelligence is the stream of reports about attacker groups, their tools and their targets. Good intelligence saves teams, and almost nobody has time to read it in full. A vendor report can run to dozens of pages of indicators, the fingerprints of an attack, such as bad file hashes, domains and addresses.

Generative AI turns that stream into something a team can act on. It summarizes a report into the points that matter for one company: does this group target our industry, do they use a tool we run, and which fingerprints should we hunt for first. Analysts ask follow-up questions in plain words instead of learning each portal layout. Junior staff gain the most. A question that once waited for the one senior analyst who knew the answer now gets a sourced draft in seconds, and the senior checks the draft instead of starting from zero.

Phishing and malware analysis support

Analysis work suits these tools because so much of it is reading. A suspicious email arrives with headers, links and an attachment. The AI walks the analyst through what it sees: where the message really came from, what the link hides, and which lines of the text push the reader to hurry. A suspicious script gets the same treatment: the analyst reads a line by line explanation of what it tries to do, then confirms the key claims against the code itself. Microsoft documentation lists this script analysis as a core use of its security assistant, and it turns an expert-only task into a guided one.

The need is real on both sides of the inbox. Our report on the warnings over AI driven phishing attacks on Gmail users shows how polished the fake messages have become. Defenders answer with the same reading skill, pointed the other way: AI that explains why a message smells wrong in words a busy employee can act on.

Secure code review and vulnerability explanation

Security work does not stop at the operations desk. The same assistants now sit inside code review, the step where new software gets checked before release. The AI reads a change alongside the human reviewer and flags patterns behind common flaws: a password stored in plain text, user input passed straight into a database query, a check that runs on screen but never on the server. It explains the risk and drafts a safer version.

The explanation matters as much as the catch. Most developers are not security specialists, and a warning they do not understand gets ignored. An AI that says why a line is dangerous, and what an attacker could do with it, teaches while it reviews. The human reviewer still decides what ships. Code holds context no model can see, and a fluent suggestion can be wrong in ways that only show up in production.

Log and query generation

Hunting through logs takes a special language. Every major platform has its own query syntax, and writing a correct query under pressure takes months to learn. Generative AI translates. The analyst types a plain request, such as showing every sign-in from a new country in the last day, and the tool drafts the query. The analyst reads it, fixes it, and runs it. Work that belonged to a few specialists opens up to the whole team.

Detection rules get the same treatment. An analyst describes a behavior worth catching, the AI drafts the rule, and the team tests it against old data before it goes live. Drafting is the cheap part of detection work. Testing decides whether a rule helps or just adds noise, and the AI does not touch it.

The other side: attackers use it too

The honest answer to how can generative AI be used in cybersecurity has a second half, and defenders who skip it plan badly. Attackers use them for the same reason: they are good at fluent text, quick research and patient drafting.

The clearest verdict comes from the UK National Cyber Security Centre assessment of AI and the cyber threat, published in January 2024. The Centre judged that AI will almost certainly increase the volume and heighten the impact of cyber attacks, and that the biggest gains land in social engineering, the craft of tricking people. Generative AI writes phishing without the spelling and translation errors that once gave it away, so a criminal working in a second language sends flawless local text. The Centre also judged that AI lowers the barrier for novice criminals and hackers for hire, which feeds the ransomware threat, since phishing so often provides the first way in.

Beyond the inbox, the same tools clone voices for fake executive calls, build faces for video fraud, and help reshape malicious code so old signatures miss it. The NCSC kept one limit in view: the most advanced uses, such as AI trained on quality exploit data to write truly new malware, stay with well resourced state actors for now. For everyone else, AI sharpens old attacks rather than inventing new ones. That is still a serious change, because higher volume and better quality break defenses tuned for the error filled lures of yesterday.

The key concern: sensitive data walks out

Asked what is a key security concern when using generative ai, security leaders give one answer before all others: sensitive data exposure. The risk is not a hacker breaking in. It is your own staff, pasting secrets into a helpful box.

Picture the daily version. An analyst drops an incident report into a public chatbot, and the report holds customer names and internal addresses. A developer pastes a block of source code that carries a live access key. A manager uploads a contract. Each paste feels private, and each one moves company data onto systems the company does not run, under terms it never read. Depending on the service, that text can be stored, reviewed by people, or used to train future models. The leak needs no attack at all, which is why the habit has a name: shadow AI, tools adopted faster than policy can follow.

The fix is boring and works. Companies approve specific tools whose contracts bar training on customer data and state how long inputs are kept. They publish a short rule on what may never be pasted, and they give staff a fast approved option, so the rule does not push the habit underground. The concern never fully disappears, because these tools grow more useful the more context you feed them. Managing that tension, rather than banning the tools, is the actual job.

Three more limits to plan around

The second limit is the confident error. Analysts call it hallucination: the model states something false in the same tone as the truth. In a security summary that can mean an invented indicator, a wrong timeline, or a cause the evidence never showed. A team that acts on it wastes hours or accuses the wrong account. The defense is procedural. Every claim in an AI draft gets traced to a log line or a document before it enters the case file, and anything untraced gets cut, however good it sounds.

The third limit is that the AI itself can be attacked. Prompt injection hides instructions inside content the AI is asked to read: a web page, an email or a file that quietly orders the assistant to ignore its task and do something else. The risk grows as assistants gain tools, because a fooled assistant with access can act. In December 2025, the NCSC warned that prompt injection may never be fully fixed, because a language model draws no firm line between instructions and data. It tops the OWASP list of risks for AI applications for the same reason. The same weakness sits behind the security risks of AI browser agents, which read untrusted pages all day. Teams design for it as the NCSC advises: limit what the assistant can touch, require approval for actions that matter, and log everything.

The last limit is accountability. When an AI summary sends a team the wrong way, the vendor does not answer for it. The analyst who signed the case does. That fact explains why mature teams keep people in the decision seat and treat autonomy as something earned task by task, starting with low risk work.

Putting it to work safely

None of these limits argues for standing still. Attackers already use the tools, and a team that refuses them reads slower than the people attacking it. The teams that gain the most follow a short set of working rules, close to how companies roll out practical AI agent use cases elsewhere: start narrow, keep approvals human, and widen only where the record stays clean. It is also why boards treat AI governance in a business context as a standing topic, not a one time purchase.

Conclusion

So, how can generative AI be used in cybersecurity? As an assistant that reads, sorts, explains and drafts at a speed no hiring plan can match: triage and incident summaries, threat reports digested into actions, phishing and scripts explained line by line, code reviewed with the risks spelled out, and queries written from plain questions. Used that way, with approved tools, checked output and people in charge of decisions, it gives scarce analysts their time back. Used carelessly, it leaks the data it was fed and believes its own fluent mistakes. The tool is the same. The discipline around it decides which story a team gets.