The AI Learning Hub
Published 1 November 2026

AI and Cybersecurity: What Changed for Attackers, Defenders, and Everyone Else

Most writing about AI and security picks one side. One version says AI hands attackers superpowers. The other says AI will finally let defenders keep up. Both are partly true, and neither tells you what to do differently on Monday morning.

This post looks at three changes: what AI gives attackers, what it gives defenders, and the new thing it adds that both sides now have to care about, which is the AI systems themselves. It ends with the habits that work against all three. Like everything in this journal it is dated, 1 November 2026. The tools will change faster than the advice.

1. What changed for attackers

AI has not invented many new kinds of attack. What it has done is make old attacks cheaper, faster and more convincing.

Phishing lost its tells. For years, the advice was to look for bad grammar and awkward phrasing. Language models removed those clues. A scam email can now be fluent, personalised and written in the target’s own language, at almost no cost per message.

Voices and faces can be faked. A short recording of someone speaking is enough to clone a voice convincingly. In a widely reported 2024 case, an employee of the engineering firm Arup in Hong Kong transferred about US$25 million after a video call in which the “colleagues”, including the chief financial officer, were deepfakes. The technique works because it targets trust, not technology.

Reconnaissance got faster. Gathering public information about a company, its staff and its systems used to take time. AI can summarise it quickly and turn it into a tailored pretext.

Writing malware got somewhat easier, mostly for people who already had some skill. AI coding help speeds up routine parts of the work. It has not, so far, turned beginners into elite attackers.

The common thread is scale. Attacks that once needed effort and skill for each target can now be run against many targets at once, with quality that used to require that effort.

2. What changed for defenders

The same technology helps the other side, mostly by taking on the reading and writing that buries security teams.

  • Triage. Security teams receive far more alerts than people can investigate. AI can summarise an alert, pull in related logs and suggest whether it looks worth a human’s time.
  • Investigation. Analysts can ask questions about their data in plain language instead of writing complex queries, and get a first draft of a timeline.
  • Explaining code and malware. AI is good at reading an unfamiliar script and explaining what it does, which speeds up analysis.
  • Writing. Incident reports, detection rules and policy drafts all start faster.

The catch is the one every AI user meets: the output sounds equally confident whether it is right or wrong. In security, a wrong “this alert is harmless” is the most expensive mistake there is. The teams getting value from AI use it to speed up human judgment, not to replace it, and they measure how often it is wrong before they trust it with anything automatic.

3. The new attack surface: AI systems themselves

The biggest genuinely new problem is that organisations are now deploying AI, and AI systems can be attacked in ways traditional software cannot.

Prompt injection is the main one. A language model can’t reliably tell the difference between instructions from its owner and instructions hidden in the content it reads. If an AI assistant reads an email, a web page or a document containing hidden instructions, it may follow them. This becomes serious when the assistant can act: send messages, call tools, or read private data.

Data leakage is the everyday version. Staff paste confidential material into public AI tools to get help with their work. In 2023, Samsung restricted staff use of public chatbots after engineers pasted internal source code into one. Most cases are less public but just as real.

Over-privileged AI agents are the growing one. An AI agent given broad access “to be helpful” becomes a valuable target, because whoever can influence it inherits its permissions.

The habits that work against all three

The encouraging part is that most of the defences are not new. They are old good practice applied consistently.

  1. Verify requests through a second channel. Any unusual request for money, credentials or data, whether by email, call or video, gets confirmed through a channel you already trust, such as calling back on a known number. A deepfake can’t answer a call it didn’t make. Some families and teams now agree a code word for this.
  2. Use phishing-resistant sign-in. Passkeys and hardware security keys can’t be phished the way passwords and SMS codes can. They remove the payoff from even a perfect phishing email.
  3. Give people an approved AI tool. Banning AI tends to push it into personal accounts. Providing an approved tool, with clear rules on what data can go into it, reduces leakage far more.
  4. Treat AI output as a draft, especially in security. Let AI speed up triage and writing, but keep a human decision on anything that closes an incident, blocks a user or changes production.
  5. Give AI systems the least access they need. Ask of every AI feature what it can read, what it can do, and whether it reads content from outsiders. If it can read private data, take in untrusted content and send information out, contain it before you ship it.

Where to go deeper

AI for Cybersecurity covers both directions in depth: using AI in the security team, and defending against AI-enabled attacks. Securing AI Systems focuses on protecting the AI you deploy, including prompt injection and agent permissions, and our threat-model canvas puts it on one page. If you are new to how these models work, AI Essentials is the place to start. All free, like everything on the AI Learning Hub.

Related free courses