AI and cyber security: the new arms race
AI is changing cyber security from both sides at once. Criminals and spies use it to write better scams and even run attacks on autopilot, while defenders use it to spot intruders in seconds and fix flaws before anyone finds them. Here's what's really happening, the companies leading the fight and what it means for you.
The new arms race
How attackers use AI
- Better phishingAI writes convincing, error-free scam emails and messages in any language, tailored to each target.ncsc.gov.uk
- Fake voices and facesCloned voices and deepfake video calls trick people into sending money or passwords.kraftintech.com
- Faster bug huntingAI helps attackers find and exploit software weaknesses faster, shrinking the time defenders have to patch.ncsc.gov.uk
- AI agents doing the hackingIn 2025 a state-backed group used an AI coding agent to carry out most of a spying campaign by itself.anthropic.com
How defenders use AI
- Spotting the unusualAI learns what normal activity looks like on a network and flags anything odd within seconds.darktrace.com
- An assistant for every analystAI summarises alerts, investigates incidents and drafts responses, so small security teams can keep up.microsoft.com
- Finding bugs before attackers doAI agents now search software for unknown weaknesses and help fix them before anyone can exploit them.blog.google
- Catching AI misuseAI companies monitor how their own tools are used and shut down accounts running attacks.anthropic.com
What's next: The UK's National Cyber Security Centre says AI will "almost certainly" make cyber attacks more frequent and more intense by 2027, and expects a growing divide between organisations that keep up with AI-enabled threats and those that don't.ncsc.gov.uk
Did you know?
- Several requests a secondAt its peak, the AI in the 2025 spying campaign made thousands of requests, often several per second, a speed no human hacker could match. It also sometimes "hallucinated" passwords that didn't work.Source: anthropic.com
- "We're security testers"The attackers got the AI to help by pretending to be a legitimate security firm testing its own defences, and by splitting the attack into small, innocent-looking tasks.Source: siliconangle.com
- A bug in the refereeIn a US government AI hacking contest, a university team's AI found flaws in SQLite, the very software that was running the competition's scoring.Source: gatech.edu
- Big Sleep stops an attackGoogle's AI agent "Big Sleep" found a dangerous flaw that only attackers knew about, and Google fixed it before it could be used: a first for AI, Google says.Source: blog.google
- More bots than peopleIn 2024, for the first time in a decade, bots made more than half (51%) of all web traffic, helped by AI making them easier to build.Source: imperva.com
- $32 billionGoogle's biggest ever purchase was a cyber security company: Wiz, which protects cloud and AI systems, bought in 2026.Source: dataconomy.com
AI itself is now a target
- Prompt injectionHidden instructions in a web page or email trick an AI assistant into leaking data or taking harmful actions.Prompt injection →
- Poisoned dataAttackers slip bad examples into the data an AI learns from, so it behaves wrongly later.Training data →
- Leaked keys and over-powered agentsAI apps built in a hurry expose secret keys, or give AI agents more access than they need.Built with AI →
The bot battle
Bots are programs that visit websites automatically. Some are useful, like search engines; others steal content, try stolen passwords, buy up tickets or flood sites. AI makes bad bots cheaper to build and better at pretending to be human.
- More bots than peopleBots made 51% of web traffic in 2024, and bad bots more than a third of it.Source: imperva.com
- AI crawlersAI companies send bots to collect web pages for training. Since July 2025 Cloudflare blocks them by default for the sites it protects, unless the owner allows them.Source: technologyreview.com
- Fighting AI with AICloudflare's AI Labyrinth sends bots that ignore "no entry" signs into a maze of AI-generated decoy pages, wasting their time.Source: blog.cloudflare.com
Running a website? Turn on the bot protection your host or CDN offers, add a "prove you are human" check to sign-up and login forms, and limit how many requests one visitor can make.
The companies leading AI security
Examples of the big names and specialists building AI into cyber defence. Not a ranking, and no one pays to be listed.
- DarktraceAI that learns what's normal for an organisation and interrupts attacks in progress within seconds.Says it can stop ransomware and phishing attacks in seconds.
- CrowdStrikeCharlotte AI, an AI security analyst in its Falcon platform that takes on time-consuming investigation work.An AI analyst that works alongside human security teams.
- MicrosoftSecurity Copilot, an AI assistant that helps security teams investigate and respond "at machine speed".Built into Microsoft's security products.
- GoogleBig Sleep, an AI that hunts for unknown software flaws, and Wiz, a cloud and AI security platform.Bought Wiz for $32 billion in 2026.
- Palo Alto NetworksAI-driven security platforms, now including identity security for people, machines and AI agents.Bought CyberArk for about $25 billion in 2026.
- SentinelOnePurple AI, an agentic security analyst that triages alerts, investigates and hunts for threats.Automates the first look at every alert, so people focus on real threats.
- DatadogBest known for monitoring software and systems; its Bits Security Analyst is an always-on AI teammate that investigates threats and security alerts.Also offers an AI agent that investigates outages (see our software delivery guide).
- CloudflareProtects websites from attacks and bad bots, using machine learning and behaviour analysis across its global network.Blocks AI crawlers by default since 2025, and traps misbehaving bots in an AI-generated maze.
- AkamaiA long-standing web security and delivery network; its Firewall for AI protects AI apps from prompt injection and data scraping.Launched Firewall for AI in April 2025.
- LakeraProtects AI applications themselves from attacks such as prompt injection. Now part of Check Point.Made the prompt-hacking games in our AI Playground.
What it means for you
For you and your family
- Use passkeys or two-step verification on email, bank and social accounts.
- Agree a family code word for urgent calls asking for money.
- Slow down on anything urgent: scammers use AI to sound real, not to be patient.
- Keep phones and computers updated; AI helps attackers exploit old bugs faster.
For small businesses
- Train staff to verify payment or password requests by phone, using a known number.
- Give everyone approved AI tools so they don't paste company data into personal ones.
- Turn on automatic updates and multi-factor login for every business account.
- Ask suppliers how they protect AI features that touch your data.
How we chose the evidence
We relied on national cyber security agencies, such as the UK’s National Cyber Security Centre, and on first-hand reports from the companies involved, such as Anthropic and Google, checked against independent news coverage. Company descriptions come from their own sites and are examples, not recommendations; no one pays to be listed. For the scams you’re most likely to meet, see how to spot AI fakes and scams, and for your data, see AI and your privacy.