AI vs AI: How Attackers and Defenders Are Both Using Artificial Intelligence in 2026

Artificial intelligence is now standard equipment on both sides of a cyberattack: criminals use it to write flawless phishing messages, clone voices and probe systems at machine speed, while security teams use it to flag abnormal behaviour and automate the first stage of incident response. For a regular internet user, the practical consequence is simple — the old advice to look for bad grammar and odd formatting no longer protects you.

This article sticks to documented behaviour rather than vendor hype: what attackers are verifiably doing with AI, what defenders are actually deploying, and which habits still hold when both sides have automation.

How Attackers Are Using AI

Defending yourself starts with the tools you use every day: see which AI assistant is actually the most private before pasting anything sensitive into one.

Three uses are well documented in public threat research. None of them are science fiction, and none require the attacker to be sophisticated.

Phishing personalisation at scale

Generative models remove the two limits that used to cap phishing quality: language fluency and research time. An attacker can scrape a target's public profile and produce a message that references a real project, a real colleague and the right internal tone. Microsoft's security researchers documented a multi-stage phishing campaign in May 2026 that used a plausible corporate "code of conduct" pretext and ended in adversary-in-the-middle session-token theft — the important detail being that the lure looked entirely routine.

Voice and video cloning

Cloning a voice from a short public sample is now trivial, and it has moved from proof-of-concept to operational use. Threat intelligence reporting in 2026 described financially motivated groups using AI-generated executive voice calls as a primary social-engineering step against banks, and the OECD's AI incident register catalogues deepfake-driven phishing against cryptocurrency targets. The attack pattern is always the same: urgency, authority, and a payment or credential request.

Automated reconnaissance and vulnerability discovery

AI assists in reading code, summarising exposed configuration and prioritising which of thousands of scanned hosts is worth attacking. It does not usually invent novel exploits — it compresses the time between a vulnerability becoming public and being used against unpatched systems.

Scale is the headline change. In June 2026, TechCrunch reported that Google sued to dismantle an alleged Chinese cybercrime network accused of using AI to defraud hundreds of thousands of victims — an indication that AI-assisted fraud is now run as industrialised infrastructure rather than one-off scams.

How Defenders Are Using AI

Defensive AI is less dramatic and, honestly, further along than most people assume. It shows up in four places.

• Behavioural anomaly detection: EDR and XDR platforms model what normal looks like for each account and device, then flag deviations — a login from an impossible location, a process spawning an unusual child, mass file access at 3am.
• Alert triage and correlation: AI clusters thousands of low-value alerts into a handful of incident narratives so analysts stop drowning in noise. This is where the measurable time savings are.
• Email and identity defence: models score message intent and sender behaviour instead of matching known-bad signatures, which is the only viable answer to AI-written phishing with no reused text.
• Automated first response: isolating a host, revoking a session token or forcing a password reset within seconds of a high-confidence detection, before a human reads the alert.

The honest limitation: defensive AI reduces time-to-detect, it does not prevent compromise. It also generates false positives, and an automated containment action that quarantines the wrong laptop has a real business cost. Every serious deployment keeps a human in the loop for irreversible actions.

What This Means for Regular People, Not Just Enterprises

You are not the target of a bespoke campaign, but you are absolutely in scope for automated ones — and the automation has removed the tells you were trained to look for.

• Perfect language is now the default. A message with no errors is not evidence of legitimacy.
• Voices are not identity proof. A call from "your bank" or "your son" that sounds right can be synthetic.
• Context is cheap to fake. Scammers can reference your employer, your recent purchase or your building's name from public data.
• Speed of exploitation has increased. Reused passwords surface in credential-stuffing attempts faster after a breach than they used to.

The upside is that your defences also improved automatically. Your email provider, browser and operating system all run AI-assisted filtering on your behalf, which is why most of this traffic never reaches you. Your job is to handle the small fraction that does. Our internet privacy guide covers reducing the public data these campaigns feed on.

How to Protect Yourself From AI-Enhanced Scams

A short checklist beats a long lecture. These are ordered by how much risk they remove per minute spent.

1. Verify through a second channel. Any urgent money or credential request — call, email, WhatsApp or video — gets confirmed on a number or account you already had. Agree a family code word for voice-based emergencies.
2. Use phishing-resistant MFA. Passkeys or a hardware key defeat adversary-in-the-middle phishing kits that happily relay SMS and app codes.
3. Make every password unique. Credential stuffing is the most automated attack that reaches ordinary people; check your habits with our password strength checker and stop reusing variants.
4. Slow down on urgency. Every AI-enhanced social-engineering script depends on you acting within minutes. Waiting ten minutes defeats most of them.
5. Keep devices patched. Automated reconnaissance finds unpatched systems fastest; enabling automatic updates removes you from the easiest bucket.
6. Protect traffic on untrusted networks. On public WiFi, a VPN closes off DNS manipulation and captive-portal tricks — see What Is a VPN? if you are unsure what it does.
7. Assume voice and video can be faked, and treat screenshots, invoices and "proof of transfer" images as unverified until confirmed at the source.

The Bottom Line

AI did not create a new class of cyberattack; it removed the friction from the old ones and simultaneously gave defenders faster detection. The result is an arms race where content quality no longer signals trustworthiness. Verification habits, phishing-resistant MFA and unique credentials are what carry you through — the same fundamentals, applied with less benefit of the doubt.