The year of the sword: AI didn’t create our security problems, but it just stopped us from ignoring them.

In Brief (TL;DR)

AI hasn't created new security problems, but it has accelerated them. From vulnerabilities being discovered and exploited faster to sensitive data finding its way into AI tools and AI agents gaining access to critical systems, organizations are facing familiar risks at a much greater speed and scale. The answer isn't to reinvent security, but to strengthen the foundations we already know work, including Zero Trust, Defense in Depth, data governance and identity management.

AI is finding vulnerabilities faster than anyone can patch them, employees are pasting confidential documents into chatbots, and organisations are putting one AI agent after another to work. Read the headlines and you would think the attacker has suddenly become unstoppable, and that the defender can only watch.

That picture is only half right. AI is sharpening the sword, that much is true. But security has always been a contest between sword and shield. Every better weapon eventually produced better armour, and every new siege engine a thicker wall. The question is not whether the attacker now has a sharper sword. It is whether we build the shield in time.

Look closely at what AI is changing in security (set aside the impact on the operations by AI deciding for you) and you find three large themes: vulnerabilities that are found and exploited faster, data that leaks into AI tools, and AI agents that are given more and more access. None of them is new. We have known them for years, and some even for decades. What has changed is the speed, and with it the urgency.

In April 2026, Anthropic launched Project Glasswing, giving a group of technology and security companies access to an AI model that can find flaws in critical software. Initiatives like this show us what AI can do. Within a month, there were more than ten thousand high- or critical-severity vulnerabilities found, many of them in code that had been running in production for years. The bottleneck was no longer finding the flaws but verifying and fixing them. That is good news for defenders and a warning at the same time. What a defender can find with AI, an attacker will eventually find too.

The window between a vulnerability becoming known and being exploited used to be measured in weeks. A handful of well-resourced groups would put it to use first, and only later would a working exploit circulate widely. That window has closed, and in many cases reversed. In M-Trends 2026, Mandiant estimates that the mean time to exploit has dropped to minus seven days: on average, attackers are using a vulnerability a week before a patch even exists. AI will only push that further.

So yes, continuous vulnerability discovery, risk-based prioritisation and fast, automated patching matter more than ever. But if your security depends on patching faster than an attacker can write an exploit, you had already lost that race, with or without AI. There will always be zero-days, legacy systems that cannot be patched and flaws nobody has found yet.

It also helps to look at what happens once an attacker is inside. According to Mandiant, attackers spent a median of 14 days in a network in 2025 before anyone noticed. For espionage campaigns the figure was 122 days, and some went undetected for more than a year. Most organisations that get breached have been compromised for some time before they find out. Whether that first step took ten minutes or ten weeks makes little difference to the outcome. What matters is whether the attacker can get in at all, what they can reach once inside, and how quickly you notice.

That is why the answer is not only faster patching, but an architecture that assumes something, somewhere, is already open. Castle builders understood this long before it had a name. They did not rely on a single wall, but on a moat, an outer wall, an inner wall and a keep, so that one breach never meant the fall of the whole castle. Defence in Depth applies the same idea to IT. Zero Trust adds that no user, device or connection is trusted because of where it sits. Segmentation ensures a compromised system is never the key to everything, least privilege limits what a hijacked account can do, and detection, increasingly supported by AI, shortens the time an attacker can stay out of sight.

Data has been leaking out of organisations for as long as there has been data to leak. Every era simply had its own channel. In the nineties it was the email to your private address. Then came the USB stick with the customer list on it, followed by business email on personal phones and files synced to a private Dropbox or backups to personal iClouds. Since 2022 it has been the confidential document pasted into a chatbot.

So, the problem is not new. What does stand out is how little we ever did about it. Data classification often remained a paper exercise, DLP tools ran in monitoring mode, and most organisations still have no clear picture of which sensitive data sits where. For a long time, we got away with that. Now we can't, for two reasons.

The first is scale. A USB stick was used now and then. AI tools are used dozens of times a day, and every prompt is a potential leak.

The second is human. People want to be smart, and to be seen as smart. In a 2018 study by Heck, Simons and Chabris, 65 per cent of Americans considered themselves more intelligent than average. Technology that makes you smarter in an instant is hard to resist, which is why generative AI can no longer be stopped. Block it, and usage does not disappear. It goes underground; into personal accounts and tools nobody controls.

So the answer is not a ban, but finally doing the homework we have been putting off for years. Know which sensitive data you have and where it lives. Classify it and enforce policies that also apply to AI tools. Know which AI services are in use and what flows into them. And above all, give people approved AI tools they actually want to use, because a safe alternative nobody uses protects nothing.

AI agents may sound like something completely new: software that carries out tasks on its own, accesses systems and makes decisions. The concern is understandable. An agent with too many privileges, or one that behaves unpredictably, can do a great deal of damage. But look at what is already running in your organisation. Service accounts have existed since the mainframe era. Since then we have added cryptographic keys and certificates, API keys, cloud identities and RPA bots. AI agents are the latest addition to a long line of non-human identities, and in many organisations those identities already far outnumber the people.

They are also regularly involved in accidents. In CyberArk's 2025 State of Machine Identity Security Report, API keys and SSL/TLS certificates were each involved in roughly a third of incidents linked to machine identities. So the problem is not that agents are new. The problem is that we never properly got machine identities under control. Service accounts without an owner, API keys that are never rotated, certificates that expire on a Friday afternoon: every security professional has seen them.

The principles for fixing this are not new either. ISO 27001 has described them for two decades, and NIS2 and DORA have since made them a legal obligation for many organisations: unique identification, an owner for every identity, least privilege, segregation of duties, periodic reviews and a record of what happened. None of these principles distinguishes between an employee, a service account or an AI agent. Organisations that already govern their machine identities can add agents with relatively little effort. Those that don't, are now receiving the bill for years of postponement, only faster and at a larger scale.

Vulnerabilities, data leakage and machine identities: three domains every security professional has known for years. AI did not invent them. It accelerated them, and in doing so exposed the shortcomings we left unaddressed for too long.

That is uncomfortable, but also reassuring. We don't need to reinvent the wheel. We know what works: Zero Trust, Defence in Depth, data governance and identity governance for humans and machines alike. We just need to finally do it properly, and faster than before.

And we don't have to do it by hand. The same technology that sharpens the sword can strengthen the shield. AI can find vulnerabilities before an attacker does, spot unusual behaviour in a sea of log data and keep track of thousands of identities no team could ever review manually. The best answer to AI in the hands of the attacker is AI in the hands of the defender.

This is the year of the sword. Whether next year becomes the year of the shield is up to us.