By Vincent Thiele, Co-Founder and COO of Modat.
Executive Summary
-
Prevention matters more than it has in years. A lot of exploitation now starts before there is even a patch to install, and patching alone cannot cover that. Cutting down what you expose is about the most reliable control we have left.
-
Assume one system will fall, and make sure the others do not follow. Picture an attacker landing on one machine. Is that a bad afternoon or the start of a full-blown crisis? Mostly, that comes down to segmentation, isolation, and identity.
-
The exploitation window is already short. ENISA’s figures for July 2024 to June 2025 put vulnerability exploitation at 21.3% of intrusion vectors. Some campaigns had a disclosure weaponised within days. According to NCSC-NL, what used to take attackers days now takes hours, and soon it will take minutes. In many cases, exploitation is done before the vulnerability is known and an advisory is published.
-
Autonomous attacks are coming sooner than predicted. The UK NCSC judged fully automated, end-to-end attacks unlikely before 2027. Yet Google Threat Intelligence Group is already watching autonomous frameworks run reconnaissance and check vulnerabilities with little human involvement, and NCSC-NL warns that AI is automating the attack chain. Our own scan data at Modat points the same way.
-
Human approval per action does not scale against machine-speed attacks. The CISO’s job shifts from approving each response to setting the guardrails that automated defence works within.
-
Data quality determines whether machine-speed defence is safe. Automated prevention and containment are only as good as the asset and ownership data underneath them. If your external inventory is out of date, a defensive action can easily turn into an outage you caused yourself.
-
Change what you report to the board. Boards love scan coverage, and it tells them next to nothing. I would rather report how quickly we go from spotting an external asset to knowing who owns it, and how much exposure we closed before anyone touched it.
Introduction
If organisations want a fair chance at protecting themselves, security practice must change quickly. Attackers are already putting autonomous agents to work, and that is spreading fast. A defence model where a person signs off on every step simply cannot keep up with that.
As a former CISO and Head of Cyber Operations, I always expected that one day it would come down to machines fighting machines, models against models, AI against AI. As a practitioner, I also spent years trying to get the basics in order: patch management, segmentation and isolation, identity and access management, and software security. The list never really ends. I should be upfront that I am biased, since I now sell a security product. But I still defend organisations on a daily basis. I will always be a practitioner.
For a long time, I argued that identification, detection, and response mattered just as much as prevention. Looking at the attacks we see today, I am no longer so sure. At least not in the way we have traditionally done detection and response. Security by obscurity is gone, and so is the idea that you can hide in huge amounts of data. My honest view is that plenty of CISOs and practitioners do not realise how exposed they are right now. At Modat, we spend our days researching two things: what agentic attackers can do, and what is sitting exposed on the internet. Put those together, and the picture is not pretty. Attacks are now so fast and so widespread that I find it hard to believe even the smallest gap in your external exposure will go unnoticed over the next year. Sooner or later, an agentic attacker, or a whole swarm of them, will find it and use it. There is a real chance that has already happened in your organisation, and if it has not, it may well happen within the next year.
So, here is where I have landed, in three points. The rest of this piece builds on them.
-
Prevention first. CISOs need to go back to security 101 and reduce exposure before anything else. Zero-days are becoming more common, so patching on its own is not enough.
-
Architecture that contains the breach. Build your environment so that one compromised system cannot take everything else down with it. That is defence in depth, designed in from the start instead of bolted on later.
-
No human in every loop. Soon there just will not be time for a person to sign off on each response. People should set the rules beforehand and let the machines get on with it.
Why this matters now
The threats themselves are not the big change. What is new is that big chunks of the attack chain now run by themselves, with nobody sitting at a keyboard.
This summer, NCSC-NL pointed out that the time between finding a flaw and exploiting it has dropped from weeks to days, and in some cases to a few minutes. It expects AI to push that down even further, and notes that AI can already automate, speed up, or even carry out parts of an attack on its own, from reconnaissance to exploitation. Back in May, it had put it more bluntly: attackers used to need days to exploit a flaw, now it is hours, and soon it will be minutes.
The UK NCSC reached a similar conclusion in its assessment of AI and cyber threats to 2027. It says the time between disclosure and exploitation has shrunk to days, and that AI will almost certainly shrink it further. It also expects a digital divide between systems that keep pace with AI-enabled threats and a large share that become more vulnerable. But in that same assessment, it judged fully automated, end-to-end advanced attacks unlikely before 2027.
From where we sit, that 2027 timeline looks optimistic. In its May 2026 AI Threat Tracker, Google Threat Intelligence Group reports agentic frameworks being used against technology firms. They switch between reconnaissance tools on their own, validate vulnerabilities, and track the attack surface with little operator input. Google also documents the first case of a criminal group using AI to find and weaponise a zero-day. In our own scan data at Modat, we see the same pattern: exposed services being probed and checked faster and more systematically than before.
Exposure is still one of the main ways in. ENISA’s Threat Landscape 2025 lists vulnerability exploitation at 21.3% of intrusion vectors, and in nearly 70% of those cases it led to an intrusion.
In real life, this looks pretty mundane. Someone spins up a new web app on a company subdomain, hosted with a cloud provider. The server’s admin login is reachable from the internet, and because it was set up outside IT, it never makes it into the asset inventory. While a quarterly review might spot it weeks later, an autonomous agent scanning the internet will find it within hours. At that speed, only the defences already in place count. That login should never have been public, and even if someone got through it, it should not have led them anywhere else.
What teams see in practice
Scans run on your calendar, attackers do not
Here is a pattern you will probably recognise. A team scans its external footprint on a fixed cycle, and whatever comes back becomes “the attack surface”. Then, between two scans, someone sets up shadow IT, a config quietly drifts, or an old DNS record is left hanging. Nobody looks at any of it until the next run. The problem is not the scan itself, but its timing. Scans follow your calendar, while your exposure changes daily. On top of that, most teams face so many alerts that fixing them all is simply not possible. What actually helps is continuous discovery, tied to who owns each asset and what depends on it, so you know what to fix first.
Hostile infrastructure moves faster than indicator lists
You get a list of indicators, block them, and by then the infrastructure has moved. Attackers set up servers, use them briefly, and dump them. Address lists go stale quickly, so you miss real activity and end up blocking harmless shared hosting. You get much further by looking at how infrastructure hangs together and behaves over time. Then automation can deal with the whole cluster instead of one IP that will be gone by tomorrow.
Back to security 101
Attackers can now find and test your exposed services day and night. However, if there is no way in, there is no incident, and that is about as cheap as security gets. Prevention is not the exciting part of the job. It is the part that is already doing its work before anyone shows up.
Of course, patching still matters. It just does not cover you anymore. Quite often the exploit is out before the fix is, so when the attack comes, there is nothing to install. The question I would ask is no longer “How fast can we patch?”, but “Why could anyone reach this in the first place?”
For me, security 101 comes down to five things:
-
Know what you expose. That means a current view of everything you have on the internet, including the things that never made it into the inventory.
-
Get rid of what should not be there. Old management interfaces, test environments nobody switched off, forgotten subdomains, legacy services. An agent will find those first.
-
Lock down what has to stay. Admin access sits behind strong identity, only certain networks can connect, and management planes are never exposed directly.
-
Give every asset an owner. Nobody fixes or restricts a service that nobody owns, and automation cannot defend it with any confidence either.
-
Build software securely. This one is easy to forget. A bug in your own code can be a zero-day too, so secure development and keeping dependencies tidy really do pay off.
Architecture that contains the breach
At some point prevention will fail. It always does. What it costs you then depends on architecture. The aim is to make sure one hacked system stays one hacked system. Easy to say, but much harder to build.
Once an autonomous attacker gets in, it keeps going. It maps the network, jumps from one system to the next, and collects privileges on the way, at a speed no person can match. A flat network, shared passwords, and service accounts with too much access make that easy. Defence in depth is there to slow it down at every step:
-
Segmentation and isolation. Your internet-facing systems should not be able to talk straight to your crown jewels. Keep production, management, and user networks apart.
-
Identity as the perimeter. Least privilege, no standing admin rights, phishing-resistant authentication, and different credentials for each environment.
-
Limit the blast radius by design. Build each system as if the one next to it has already been hacked.
-
Containment you can undo, built in ahead of time. Put the isolation points into the design so a machine can shut them in seconds. No change board needed.
Do not underestimate that last one. Imagine containment that requires someone to write a firewall change at 03:00. You will be too slow, and it was your architecture, not your team, that made it so.
Rethinking detection and response
Can automated response act without someone signing off? CISOs and heads of cyber operations have argued about that for years. To me, that debate is pretty much settled now. Attackers use AI to speed up reconnaissance, vulnerability research, and exploitation, and nobody on their side goes home at night. An analyst who has to triage, escalate, and then wait for approval is simply too slow when the window is minutes wide.
So the questions change. Will your prevention hold up? And when automation steps in, will it hit the right target? Automated containment can only be as good as the data behind it. If you are working from an incomplete or outdated picture of your external exposure, your own defensive action might open new holes or take down something you needed.
Keeping a person in every loop is not the answer. Instead, move them from deciding on each action to deciding the rules ahead of time:
-
Before an incident, people decide which assets automation is allowed to touch, which actions it can take, how confident it needs to be for each one, and the maximum blast radius.
-
During an incident, machines act within those rules at machine speed and log every decision they make.
-
After an incident, people review those decisions, fix ownership records, and tighten the rules.
Some people call this human-on-the-loop instead of human-in-the-loop. Whatever you call it, it only works when your asset, ownership, and exposure data are solid enough for a machine to make decisions with. In the end, data quality is what holds you back, much more than tooling.
Automation amplifies bad data
Automation starts with enrichment, then prioritisation, and ends with containment. Each step amplifies whatever is wrong in the underlying data. A false positive that used to cost an analyst ten minutes can now cut off a route to a payment service. Taking people out of each decision is unavoidable. Putting up with poor data is not. Agents need infrastructure context that is clean, structured, and current, with confidence levels an owner can stand behind.
Trade-offs
This approach has costs too. False positives are still there; they just turn up in other places. Better context means analysts and their agents chase fewer leads that go nowhere. Even so, before containment runs without anyone watching, decide how certain it must be before it acts and how far its actions can reach.
-
A machine making the call does not let anyone off the hook. Every rule needs a person’s name on it, especially when an action could take down a business service, affect a third party, or end up in an incident report.
-
Before any of this, do the groundwork. If the asset register cannot be trusted, ownership records are not kept up to date, or enrichment never reaches the SIEM or intelligence platform, automation should wait. Automating on top of those gaps only accelerates existing problems.
-
Scanning the internet only shows you the outside. Internal segments, misused identities, and insiders need different controls, which is another reason architecture matters.
-
If undoing an action takes a change board, a vendor call, or a press statement, a machine should not be doing it. Do not solve that by putting a person back in the loop. Change the action so it is quick to reverse.
Five decisions for the CISO
Why the CISO and not the SOC? Because each of these sets a rule your machines will end up following.
| Decision | Required context | Owner | Risk of delay |
|---|---|---|---|
| Which exposure is removed or restricted first | Attackability rather than presence, business criticality, threat context tied to the asset | CISO with exposure management lead | Agents find and use the exposure before the remediation queue reaches it |
| Where segmentation and isolation must stop lateral movement | Crown-jewel systems, trust relationships, identity paths between zones | CISO with enterprise architecture | One foothold becomes a full compromise |
| Which containment actions automation may take without approval | Confirmed ownership, confidence threshold, maximum blast radius, reversibility | CISO with the service owner | Response arrives after the attacker has moved on |
| Which hostile infrastructure is blocked before a campaign lands | Infrastructure relationships and a stated confidence level | CISO with Head of CTI | Blocklists that age out before the campaign starts, or over-blocking |
| What external exposure is reported to the board | Discovery-to-ownership timing, count of unowned assets, exposure closed before attack | CISO | Governance built on scan cycles rather than current exposure |
So what does that mean? Prevention and architecture buy you the room that automated response needs to be safe. And if assets have no owner, that is a governance problem no platform can fix. Sort out ownership first, then spend money on automation.
How it works in practice
The same workflow works for your own exposure and for the attacker’s infrastructure. People design it. Machines run it.
-
Discover. Keep a live map of everything that can be found from the outside, including assets that are not in your inventory. You will need hosts, certificates, passive DNS, service fingerprints, and history to do it.
-
Validate. Confirm ownership, exposure state, and relevance, and record how confident you are. This step decides what automation is allowed to touch later on.
-
Prevent. If there is no business reason for something to be exposed, take it down or lock it down before somebody attacks it.
-
Prioritise. Weigh technical exposure, business criticality, and threat context together. Whether something can actually be attacked matters more than whether it is there.
-
Act. Automation takes the pre-approved action, within its confidence threshold and blast radius. Anything outside those rules goes to a named person.
-
Review. Record the decision, correct the ownership record, and adjust rules, detections, and playbooks. That is what makes the next round faster.
Your next 90 days
-
Start by comparing your own asset inventory with what anyone can discover about you from the outside. Whatever does not match goes on a backlog, with a deadline.
-
Make sure every internet-facing domain, certificate, host, application, and cloud service has a named owner. Do the ones behind critical business processes first.
-
Take down or lock down exposed management interfaces, test environments, and anything else that is reachable for no good business reason.
-
Trace how someone could get from an internet-facing system to your crown jewels. Close off the routes where one foothold could turn into a full compromise.
-
Pick the containment actions automation may take without asking. For each one, decide how sure it must be and how much it may touch, and test that you can roll it back. Practise. Then turn them on.
-
Take scan coverage out of the board report. Add how long it takes to find an external asset and confirm its owner, and how much exposure was closed before anyone attacked it.
Where Modat fits
As I said at the start, I am biased here, since I work at Modat, a European Internet Intelligence company. So read this part with that in mind.
Almost everything in this piece depends on one thing: an accurate, current picture of what you have on the internet and of the infrastructure aimed at you. That is the part Modat works on. Magnify continuously scans and enriches internet-connected infrastructure. It identifies what each device actually is, using more than 15,000 device profiles, and it keeps the history, so you can see how infrastructure changes over time instead of a single snapshot. It will not segment your network or fix identity for you. It gives the people and agents doing that work better data to start from.
Here is what that looks like. An analyst, or an agent working for them, spots a suspicious domain targeting companies in their sector. They follow it to the servers and certificates connected to it, and check whether it all belongs to one operation. Then they compare it with their own organisation’s footprint. If it matches an asset nobody owns, it is no longer a research question. It becomes an exposure to close, and someone must take ownership.
Modat works inside the tools teams already use, through our API and standards-based MCP access, so analysts and their agents see the same data. People decide, and agents deliver.
Final thoughts
I spent years believing that if we could just see the attack, we could stop it. I do not believe that anymore. Not because detection stopped mattering, but because by the time a person saw the alert, read it, and asked for approval, an autonomous attacker had already moved on.
So here is the uncomfortable question I would ask every CISO reading this: if your response plan still needs someone to approve a containment at 03:00, who do you think wins that race?
The organisations that come through this will not be the ones with the most tools or the biggest SOC. They will be the ones that went back to basics, built walls that hold when one system falls, and decided in advance what their machines are allowed to do.
Machines will fight machines. That part is already decided. What is still up to you is whether yours are working from better data, behind better walls, and with clearer rules.

