Most recently, an Iran-backed group manipulatedOpens a new window OT devices to disrupt several critical infrastructure sectors — government services, energy, water, and wastewater.
The advisory warns that poor network segmentation lets AI-generated malware move laterally and hijack machines with minimal resistance.
The long-standing gaps in OT security — legacy systems on outdated software, little to no visibility over OT devices — make that worse, not better.
Frameworks from CISA and NIST already recognize network segmentation as a baseline security requirement, and industrial enterprises have been responding.
“Flat or minimally segmented OT networks remain common in the US and globally.
You probably have a PLC (programmable logic controller) somewhere on your factory floor that’s older than the intern who set up your help desk portal. It works fine. It’s been working fine for years. Nobody touches it, nobody patches it, and until last month, nobody really worried about it.
Here’s why that’s changing. Five US agencies – NSA, FBI, CISA, DOE, and EPA — just put out a joint cybersecurity advisoryOpens a new window about an active threat targeting PLCs, the industrial computers that automate machinery, manage production lines, and control plants. The attackers aren’t doing this by hand anymore. They’re using AI-generated Python scripts to skip the hard parts – complex network discovery, mimicking legitimate tools to dodge detection — and going straight after poorly segmented or vulnerable PLCs.
And those PLCs sitting on a flat, unsegmented network? That’s the part that should grab your attention.
READ MORE: IT Job Watch: Network security analyst
The old playbook doesn’t work when the attacker shows up at machine speed.
State-backed threat actors have been playing cat and mouse with US companies for years. Most recently, an Iran-backed group manipulatedOpens a new window OT devices to disrupt several critical infrastructure sectors — government services, energy, water, and wastewater. That was bad enough. Now they’re layering AI on top.
The advisory warns that poor network segmentation lets AI-generated malware move laterally and hijack machines with minimal resistance. The long-standing gaps in OT security — legacy systems on outdated software, little to no visibility over OT devices — make that worse, not better.
Dragos confirmed as much in its 2026 OT cybersecurity reportOpens a new window : only 30% of organizations have any visibility over their OT networks, while 88% struggle with detection and response. The firm also flagged that many attackers have already reached stage two — carrying out reconnaissance and testing inside the OT environment, learning control loops, and preparing for future manipulation. They’re not at the door anymore. They’re inside, taking notes.
Understanding these shifting dynamics is essential to staying ahead of rapid, machine-speed exploits before critical infrastructure is compromised.
Why flat networks were fine until they suddenly weren’t
OT systems were designed for high uptime, real-time control, and ease of maintenance. That design philosophy gave us flat, unsegmented networks where every device shares a single VLAN. It made sense at the time. Many of these networks were also air-gapped — physically isolated, with no connection to anything outside.
READ MORE: Is the traditional tech interview making a comeback?
Then digital transformation and AI projects happened. Those air-gapped networks got connected, often without additional controls bolted on. The isolation that was the entire security model disappeared, and what’s left is a network where compromising one device can give you the rest.
Trout Software warnsOpens a new window that in a flat network, an attacker who compromises a single device gains access to PLCs, HMIs (human-machine interfaces), and safety controllers sitting on the same network. A network loop or a misconfigured device can also trigger a surge in broadcast traffic, leading to congestion.
In IT, that means emails arrive late. In OT, it’s a different story — industrial processes rely on real-time feedback loops to calculate the next physical adjustment, so congestion can cause control instability. Put plainly, your machinery malfunctions or behaves dangerously.
What segmentation actually buys you
This approach isn’t novel. Divide your network into isolated subnets, placing each device in its own segment. Route all traffic between segments through a firewall configured to examine industrial protocols, permitting legitimate control commands while rejecting harmful ones. Network segmentation halts lateral movement, shields critical devices in separate segments from unauthorized access, and generates a complete audit trail because every inter-segment connection passes through a monitored logging checkpoint.
Frameworks from CISA and NIST already recognize network segmentation as a baseline security requirement, and industrial enterprises have been responding. Fortinet’s 2026 State of OT and Cybersecurity reportOpens a new window shows the share of companies with full visibility over their IT systems improved from 5% in 2025 to 14% in 2026. Forty percent said their ICS systems are less than five years old, up from 20% in 2025 — a real shift toward modernization.
But the same report found that many organizations are still establishing the fundamentals: asset visibility, network segmentation, secure remote access, and monitoring. They’re not there yet.
“Flat or minimally segmented OT networks remain common in the US and globally. OT cybersecurity often begins with gaining visibility into networks and devices, quickly followed by segmentation; however, some industrial and critical infrastructure operators have yet to start on these initial basic steps,” Richard Springer, Senior Director of OT Product Marketing at Fortinet, told Spiceworks via email.
AI didn’t invent the attack — it removed the skill barrier
Attackers are increasingly using gen AI tools like GhostGPT, HackerGPT, and FraudGPT to automate the tedious parts: writing basic scripts and ransomware, scanning networks, and analyzing large codebases for critical vulnerabilities. The tools do in minutes what used to take a knowledgeable attacker days.
Palo Alto Networks’ 2026 Global Incident Response reportOpens a new window shows that agentic AI can automate the entire attack chain — reconnaissance, exploitation, lateral movement — in minutes. In a controlled test by their Unit 42 team, a full-scale ransomware attack with data exfiltration was carried out in just 25 minutes after initial compromise.
Springer also warns that AI-enabled attacks help adversaries quickly identify assets, understand OT networks, and determine which known vulnerabilities or weaknesses are exploitable. That’s especially concerning in plants where PLCs, HMIs, SCADA systems, and other specialized devices may be older, may not have patches, or can’t be patched easily because of operational requirements.
“While these environments also rely on specialized devices with specific communication protocols, AI-driven attacks can rapidly translate and provide guidance on manipulating these devices. Previously, this would require an OT-knowledgeable attacker to have the requisite knowledge to alter the device. Most don’t, so most didn’t bother with OT. AI-enabled attacks are changing this paradigm,” he adds.
That’s the shift in one sentence: the attacker no longer needs to be an OT expert. The AI handles that part.
What to do about it — before the next advisory lands
Network segmentation remains a critical defense and will slow attackers down, but it may not be enough against AI-driven reconnaissance that can rapidly identify and exploit segmentation gaps. Experts point to ZTNA (zero trust network access) as the next layer — moving from defending static network boundaries to dynamic, identity-based ones.
“ZTNA principles in OT environments can create significant barriers for bad actors. It can add multiple layers of protection that make it significantly harder for a bad actor to move through an OT environment. Client- or browser-based secure remote access, combined with robust password management and MFA, can prevent or deter most common attack paths,” said Springer. However, he adds that because OT systems often use older, unique devices, applying ZTNA comprehensively can be difficult and must be done selectively.
The threat landscape has shifted dramatically with the rollout of gen AI and agentic AI tools. New models like Anthropic’s Claude Mythos and OpenAI’s GPT-6 Astra are believed to be capable of finding thousands of previously unknown critical vulnerabilities. OT infrastructure designed decades ago faces a higher risk than ever. But with frameworks like ZTNA, critical infrastructure providers can still slow attackers and limit the damage — and right now, slowing them down might be the most realistic goal on the table.
