What NIS2 and the EU AI Act Expect of Critical Infrastructure Operators in Ireland

Most security architecture rests on an assumption that nobody says out loud: that the thing arriving at the perimeter has been seen somewhere before. A signature exists. An IP carries a reputation. A payload matches something in a database maintained by somebody whose job is to maintain it. The whole apparatus of blocklists, feeds and categories depends on that being true often enough to matter.

Nothreat's ThreatScape analysis of over 16 million automated attacks suggests the assumption is weakening. The headline figure is a rise of nearly 470% in payloads that had never been seen before, concentrated in the final quarter of the period studied. Novel payloads went from just over 1% of observed activity in October to nearly 6.5% by December. The shape of that curve matters more than the percentage.

What the data says about industrial infrastructure

For anyone responsible for an operational estate, the sector breakdown is the part worth reading twice.

The energy sector saw a 74% increase in new payloads and malicious URLs compared with generic IoT, which points to attacks written for specific industrial targets rather than opportunistic scanning that happens to land on a PLC. Alongside it, human-driven attacks against energy IIoT rose 87%. Automated scanning is cheap and indiscriminate. A person sitting at a keyboard working through an industrial control environment is a different proposition, and it usually means somebody has decided the target is worth the hours.

The public sector figures point the same way from a different angle. Country-specific attacks made up 66% of activity against public sector and government targets, far above other sectors, while payloads never previously seen accounted for a smaller share than in healthcare or banking. Nothreat reads that as a preference for stealth and persistence rather than a lower level of threat: advanced persistent activity designed to sit inside an environment and exfiltrate slowly, often arriving through a supply chain rather than the front door.

Two further figures are worth carrying into any conversation about access. Credential stuffing attacks against VPNs went from a rounding error to 21% of all observed breaches by December, an increase Nothreat measures at 4,100 times. Around 15% of those involved manual effort using credentials already leaked on the dark web. Advanced SQL injection using SLEEP and DELAY commands to evade detection rose fourteenfold.

The problem with clean infrastructure

The case studies in the report make the point more sharply than the percentages do.

One adversary ran 2,858 advanced attacks and nearly 7,000 malicious requests from a Hong Kong IP address that appeared nowhere in global threat intelligence feeds. No alerts, no abuse reports, nothing on AbuseIPDB. Another executed over 40,000 attack requests against a client’s critical service infrastructure from an address that had never been flagged by any major threat database, with only port 22 open and no history of compromise. A third distributed activity across 226 addresses from a single operator block, twenty attempts per cycle, repeating every few days, deliberately paced to look like nothing at all.

None of those would be stopped by reputation. All three suggest adversaries buying or building infrastructure that is clean precisely because clean infrastructure defeats the defence most organisations rely on.

Why this lands harder on operational estates

An IT network can absorb this reasonably well. Endpoints get patched, agents get deployed, feeds update several times a day and somebody notices when they stop.

Operational estates work differently. Controllers, RTUs and edge gateways frequently cannot host an agent, run protocols the security tooling does not speak, and sit behind segmentation designed to keep traffic out. That segmentation does its job on threat intelligence updates as effectively as it does on attackers. The equipment carrying the operational risk is often the equipment that hears least about the threat and taking it offline to change that has consequences measured in stopped processes rather than inconvenience.

Which is a difficult position to be in when the questions start arriving.

The questions are arriving

In conversations with operators across Ireland this year, the pressure is showing up commercially well ahead of any regulator. Client security questionnaires, insurance renewals and third party risk assessments are all converging on the same question: what monitors the operational network, and what evidence exists that it works?

NIS2 sharpened that considerably by putting accountability with management bodies personally rather than leaving it as an engineering matter. Ireland has still not enacted the National Cyber Security Bill, and the European Commission referred the delay to the Court of Justice in July, though the legislative gap has done nothing to slow the commercial pressure. A manufacturer in scope sends a sixty-question assessment to its logistics provider. None of those parties have enforcement powers, and every one of them can stop a deal.

The EU AI Act adds a second layer, and its timeline shifted over the summer. The Digital Omnibus on AI came into force in late July, deferring high-risk obligations for standalone Annex III systems to December 2027 and for AI embedded in regulated products to August 2028. Annex III covers AI used as a safety component in the management and operation of critical infrastructure, so a good deal of preparation across European utilities and transport operators has been given sixteen extra months. The security obligations under NIS2 moved not at all.

What a different approach looks like

If most of what arrives is novel, and if the infrastructure it arrives from is deliberately clean, then detection must reason from behaviour rather than identity. That is a well understood idea. The harder part is doing it on equipment with almost no compute available and no route for updates.

Nothreat’s approach is a model that learns locally and continuously, acquiring new knowledge in minutes, and running on ordinary CPUs without requiring a GPU. Nothreat AIoT Defender operates as a software firewall inside edge devices in roughly 2MB of RAM, reaching 97.3% accuracy against zero-day threats within twelve minutes of starting from nothing. Deception fills the other half. Nothreat CyberEcho creates digital twins of the client's IT services, deployed as decoys to gather attack signals isolated from legitimate traffic, which produces the clean data that behavioural detection needs in order to be accurate.

CyberEcho deserves a closer look, because it answers a question that behavioural detection on its own cannot. A behavioural model scores activity on probability. A decoy sits where no legitimate user or process ever goes, so any interaction with it is by definition an attack. That is a structural fact rather than a confidence score, and it changes what an operations team has to do with the alert. There is no triage, no false positive to investigate, no judgement call at 3am. The alert is the confirmation.

For estates where every change goes through a maintenance window and a change board, deployment matters as much as capability. Nothreat integrates with existing firewalls, SIEM and EDR tooling, including Cisco, Palo Alto, Fortinet and over twenty others, with a single line of configuration. The platform deploys in minutes and sits alongside what is already running rather than replacing it. For an environment where nobody wants to touch the stack, that removes the main objection before it is raised.

There is a question that tends to arrive early in any conversation with Irish and European infrastructure operators: where does the data go? Nothreat's decoys are isolated from production systems and never collect data from the live environment. Deployment is available on-premise or in sovereign cloud, which gives operators control over residency without a trade-off against detection capability.

The strongest proof point for detection that learns rather than waits is a recent one. In mid-2025, Nothreat detected and blocked a SharePoint zero-day, subsequently tracked as CVE-2025-53770 and CVE-2025-53771, forty days before the CVE was published. The campaign, which researchers later named ToolShell, was independently documented by Microsoft, CISA and Palo Alto's Unit 42. The platform caught it because it had never seen the behaviour before and recognised it as anomalous, which is precisely the scenario that signature-based detection cannot address until somebody else gets compromised first.

Whether that specific architecture suits a given estate is a question for the engineers who run it. The principle underneath it is harder to argue with. When the majority of what reaches an industrial network has no prior record anywhere, a defence that waits to be told what to look for is waiting for information that may never come.

Join us on Wednesday 23rd September

Nothreat joins Renaissance at 11am on Wednesday 23rd September for a session on what NIS2 and the AI Act now expect of Ireland’s critical infrastructure, and what can realistically be deployed on estates full of equipment that cannot be patched or replaced. The session covers the ThreatScape findings in more detail, how detection that learns on the device works in practice, and what evidence it produces when a regulator, an insurer or a customer’s risk team asks how the operational network is monitored. Register here.