AI in Cyber Security: Where AI Helps Defenders Keep Up

Tilt-shift miniature of a busy railway sorting yard at dusk, with dozens of freight wagons arriving on many tracks while a small maintenance crew works on one wagon — Monthly vulnerability disclosures roughly doubled between January and August 2026. Most patch processes were built for the old pace.

Last week a pair of PHP flaws took government websites offline across Europe, a Node.js flaw followed the same day, and Cloudflare’s chief executive told Reuters he expects a couple of years where “every single week it’s going to feel like the sky is falling”. The reason is AI. It is now finding software vulnerabilities faster than at any point before, and it is doing it for attackers and defenders at the same time.

The usual response is that businesses should start using AI for defence. That is right, but it skips a step. In most of the information security systems I audit, the bottleneck is not spotting a vulnerability. It is knowing which systems you run, and then patching them in days rather than weeks. AI can help with both, and this article covers where it genuinely does, with real examples, and where it does not.

Short answer

What changed: AI has roughly doubled the number of software flaws found in a year, and exploits that used to take weeks can now be built in hours.
What ASD says: identify every internet-facing service daily, patch vulnerabilities of all severities in them quickly, and accept the occasional brief outage that faster patching brings.
Where AI helps defenders: finding exposed systems, deciding what to patch first, triaging phishing and alerts, spotting unusual behaviour, and checking code. Examples are in the table below.
The catch: AI security tools are suppliers and AI systems in their own right. They sit on top of your asset register and patch process. They do not replace them.

What has actually changed

The numbers are moving quickly, and they come from the people watching most closely:

Twice the flaws. Research from VulnCheck, reported by Reuters on 26 October, found AI has helped roughly double the number of software flaws found compared with a year earlier.
Monthly disclosures doubled in 2026. According to Google’s Threat Intelligence Group, monthly vulnerability disclosures rose from 5,045 in January to a record 10,740 in August.
The AI-found flaws are the dangerous kind. Of the vulnerabilities Google considers likely to have been found by AI, half allowed remote code execution, against 26% of the rest.
Less time to react. Microsoft’s 2026 Digital Defense Report puts the median time from a vulnerability being found in the wild to being weaponised at well under 24 hours. The same report says large organisations can take 30 to 60 days to fix critical vulnerabilities in their external systems.

Not every new flaw is exploited. Google counted 141 newly disclosed vulnerabilities exploited between January and August 2026, against 127 in all of 2025, which is roughly one in every 430 disclosures. The problem is that nobody knows in advance which one it will be, and the window to fix it is shrinking.

Update, 28 October 2026: ASD’s Annual Cyber Threat Report 2025-26, published on 26 October, describes AI as the defining force in the threat landscape over the past year and reports a 30% rise in published vulnerabilities (CVEs). The figures for Australian businesses are in what ASD’s 2025-26 threat report shows.

Why most patch processes were built for the old pace

Many businesses still patch on a monthly cycle, and many rely on an IT provider who does the same. That worked when an exploit took weeks to appear. It does not work when an exploit can appear within hours of the fix.

Mathspace is the example I keep coming back to. The vendor published the patch on 6 August, the breach happened on the 10th, and the patch was applied on the 29th. Nothing in that sequence needed AI. It needed a process that treated an internet-facing fix as urgent. Our vulnerability management article walks through it, and the PHP2Shell alert last week is the same pattern on a much wider scale.

What ASD says to do

In July, ASD’s Australian Cyber Security Centre published guidance on defending against AI-enabled cyber attacks. It is written for government, critical infrastructure and large organisations, but the immediate actions apply to any business with a website, a portal or remote access:

Name an executive who is accountable for the risks from attackers’ use of AI and from your own use of it.
Raise the board’s tolerance for brief outages caused by faster patching, more frequent patch windows and the occasional bad patch.
Patch or mitigate vulnerabilities of all severities in internet-facing services quickly, using automation where a bad patch is low risk or easy to reverse.
Identify every internet-facing service daily, remove the ones you do not need, and replace the ones the vendor no longer supports.
Centralise security logs and analyse them promptly, using automated triage and narrow, reversible automated responses.

Using AI is in the guidance, but it is a medium-term action, and it comes with conditions: AI to find and confirm vulnerabilities in code, and to help analyse logs and support testing, under human supervision. The basics come first. The joint guidance ASD published with its Canadian, UK and New Zealand counterparts, Opportunities for AI in cyber defence, makes the same point: strong cyber security fundamentals remain the most important thing.

Our article on ASD Essentials replacing the Essential Eight covers where the baseline controls are heading.

Where AI genuinely helps defenders, with examples

AI earns its place in defence on the jobs people never get to: watching the perimeter every day, sorting thousands of alerts, and reading every suspicious email. These are the six areas where it is doing real work now.

Defensive job	What the AI does	Examples
Finding what you have exposed to the internet	Discovers websites, portals, remote access and forgotten test systems from the outside, the way an attacker would, and flags changes.	Microsoft Defender External Attack Surface Management; Tenable attack surface tools
Deciding what to patch first	Scores each vulnerability on the likelihood of real exploitation, not just its severity rating, so the risky few go first.	Tenable ExposureAI; the free EPSS score from FIRST, a machine-learning model; the CISA Known Exploited Vulnerabilities list
Email and phishing	Reads reported and suspicious emails, decides which are malicious and explains why, so a person reviews the verdict instead of every message.	Microsoft Defender Phishing Triage Agent, part of Security Copilot, now included with Microsoft 365 E5
Endpoint alerts and investigation	Pulls together alerts from laptops and servers, writes up what happened and can contain a device automatically.	Microsoft Defender for Business automated investigation and response (in Microsoft 365 Business Premium); CrowdStrike Charlotte AI; SentinelOne Purple AI
Unusual behaviour on the network	Learns what normal looks like for your users and devices and flags or blocks activity that does not fit.	Darktrace
Your own code	Finds vulnerabilities in code you write or commission and proposes a fix for a developer to approve.	GitHub Copilot Autofix; Google CodeMender (public preview since July 2026, with a person approving every patch)

Examples only, not recommendations. Streamline does not sell or resell any of these products. Features and licensing change often, so check the current position with the vendor or your IT provider.

For most small and mid-sized businesses, the realistic way to get these capabilities is through the tools you already license, or through your IT provider or a managed detection and response service. Microsoft 365 Business Premium, for example, already includes automated investigation and response, and many organisations have never checked whether it is switched on. Before buying anything new, it is worth asking what you already pay for and what your provider already runs.

The catch: AI security tools are suppliers and AI systems too

An AI security tool usually needs broad access to do its job: your mailboxes, your endpoints, your logs, sometimes your code. That makes it one of the most privileged suppliers you have. The same questions apply as to any critical supplier, plus a few that are specific to AI:

What data does it see, where is that data processed, and is it used to train the vendor’s models?
What can it change on its own, and which actions need a person to approve them?
How would you know if it got a decision wrong, and how quickly could you reverse it?
Who in your organisation owns it, and does it appear in your asset register and supplier register?

ISO 42001 is built for exactly this: managing the AI systems an organisation uses, not only the ones it builds. Our articles on letting an AI agent loose and assessing AI tools together rather than one at a time go further.

What this looks like in ISO 27001 terms

None of this needs a new framework. It needs the controls most organisations already claim to have, working at the new pace:

A.5.9 Inventory of information and other associated assets: does the register include every website, portal, cloud service and remote access point, including the ones a supplier runs for you?
A.8.8 Management of technical vulnerabilities: is there a written timeframe for internet-facing fixes, and is it met? Can you show the dates?
A.5.7 Threat intelligence: when an alert like PHP2Shell lands, does it lead to an action with an owner, or just an email?
A.8.16 Monitoring activities: are logs reviewed promptly, and is it clear who responds out of hours?
A.5.19 to A.5.22 Supplier relationships: does your IT provider’s agreement say how fast they patch, and do you check?

When I audit A.8.8, I pick one recent critical vulnerability and ask for the dates: when the fix was released, when the organisation knew, and when it was applied. That one trail says more about readiness for AI-speed attacks than any tool list.

Questions to ask your IT provider this week

How do you find every internet-facing system we have, and how often do you check?
What is your target time to patch a critical internet-facing vulnerability, and what was the actual time for PHP2Shell?
Which security tools are you running for us that use AI, and what can they do without a person approving it?
Are the automated investigation and response features in our existing licences switched on?
If a critical fix needs a short outage, who can approve it, and how quickly?

If the answers are vague, that is the gap to close first. Our article on how to ask your IT provider the right question has more on that conversation.

AI in cyber security: FAQs

Is AI really making cyber attacks faster?

Yes. Google’s threat intelligence group reports monthly vulnerability disclosures roughly doubled between January and August 2026, and Microsoft puts the median time from discovery to a working exploit at well under 24 hours. ASD’s July 2026 guidance warns that exploits can now be developed within hours.

Should a small business buy AI security tools?

Start with what you already have. Many Microsoft 365 plans include automated investigation and response, and many IT providers already run AI-assisted tools. An accurate asset register and a fast patch process will do more than a new product sitting on top of a slow one.

How fast should we patch internet-facing systems?

At maturity level one, the Essential Eight expects fixes for internet-facing services within 48 hours where the vendor rates the vulnerability critical or a working exploit exists, and within two weeks otherwise. ASD’s 2026 guidance on AI-enabled attacks goes further, asking organisations to patch vulnerabilities of all severities in internet-facing services quickly, and to automate where it is safe to do so.

Does ISO 27001 cover AI security tools?

ISO 27001 covers them as assets and suppliers, through controls such as A.5.9, A.5.19 and A.8.16. ISO 42001 adds the AI-specific side: how decisions are made, overseen and corrected. Many organisations use both together.

How Streamline can help

Streamline helps organisations build ISO 27001 information security management systems with asset registers that include every internet-facing service, supplier arrangements that set out who patches what and how fast, and a vulnerability management process that keeps up. Where AI tools are part of the picture, our ISO 42001 consultants help bring them under the same control. If you would like to talk through how your organisation would handle the next critical alert, get in touch.

Sources: ASD’s ACSC, Defending against AI-enabled cyber attacks (6 July 2026); ASD’s ACSC, Using AI to strengthen cyber defence; Reuters, via Carrier Management (26 October 2026); CybersecAsia on Google Threat Intelligence Group analysis (6 October 2026); Microsoft Digital Defense Report 2026; Microsoft, Security Copilot with Microsoft 365 E5.

General guidance only. This article is general information, not legal, financial, safety or compliance advice, and it does not take account of your specific circumstances. Streamline ISO Consultants are ISO management-system consultants, not lawyers or licensed advisers. Standards, laws and regulator guidance change, and details were correct only at the time of writing. Always seek professional advice before acting. See our full Disclaimer.