In July 2024, a faulty content update to a widely deployed endpoint protection product crashed an estimated 8.5 million Windows devices and set off IT outages across industries worldwide.
For many operational technology (OT) operators, the incident reinforced a longstanding caution about putting software on assets that run physical processes. We’ve written before about why that caution is warranted when the software is an IT agent built for enterprise environments. Many teams are cautious even about endpoint tools designed for OT, and that caution sets the right bar. Endpoint security for OT must add protection without disrupting the process it’s there to protect.
Network monitoring is foundational to OT security because it can see assets and their communications without installing software on them. Endpoint visibility adds a view of what happens on the machine itself, like a user logging in to an engineering workstation, a process launching on a human machine interface (HMI), a file changing or a USB device typing commands into a host.
That combined view matters more as frontier AI helps attackers find and exploit vulnerabilities faster. For teams that set endpoint visibility aside, it’s a reason to take another look. As attack timelines compress, activity on a single host can be the first sign something is wrong, and spotting it early gives teams more time to respond well.
But endpoint tools developed for enterprise IT start from different assumptions than OT does. In IT, the first priority is usually protecting data, so isolating a host, restarting a machine or applying an update quickly is often the right call. In OT and other cyber-physical systems (CPS), where software monitors and controls a physical process, the first priority is keeping that process running safely. Those same actions can interrupt production, affect equipment, and in some environments, create safety risks. That’s why each one has to account for the process the asset supports, how long that asset can be unavailable, and how changes are normally planned.
According to Gartner®, “Endpoint protection (EPP/EDR) deployment in a CPS environment is an operational and engineering risk decision, not a routine IT task.”
When an endpoint tool detects something, it can do one of three things: act on its own, ask a person to approve, or alert the team. None of them is the right answer everywhere.
Which one fits depends on the asset, so it needs operations at the table. For OT teams, the goal isn’t simply to add endpoint detection. It’s to add host-level visibility without giving up the operational control, context, and change discipline these environments require. That means deciding where the technology adds value, how its actions could affect the process, and which of those three responses belongs on each asset. The five questions below offer a practical way to work through those decisions.
1. Where Could Endpoint Visibility Add Value in Your Environment?
Network monitoring keeps the wider environment in view, including controllers and devices where installing software is often impractical or unsupported, and any managed or unmanaged asset that’s easy to lose track of. What it can miss is activity that stays on the machines people use every day, like the engineering workstation where a contractor logs in to update controller logic, or the HMI where an operator plugs in a USB drive to move a file.
Those machines are a good place to start with endpoint visibility. A trusted account used in an unusual way, or a USB device that types commands instead of storing files may never show up on the network, or may look like routine work when it does. On the host, both leave a trail. We’ve written about why insider threats in OT call for endpoint sensors alongside network monitoring for exactly this reason.
2. What Could an Endpoint Tool Mean for the Process?
Start with what the tool would actually do. Enterprise endpoint detection and response (EDR) platforms are mature, and much of their value comes from automated response. When they detect a threat, they can quarantine a file, terminate a process or isolate the host from the network without waiting for an analyst. In enterprise IT, that speed is a strength.
Now put those same actions on an operational asset. If a detection quarantines a component of HMI runtime software during a shift, operators can lose their view of the process. If an isolation policy disconnects an engineering workstation without exceptions for industrial traffic, the workstation also loses its connection to the controllers it manages. Neither scenario requires the asset to be compromised. It only takes the tool acting on a detection, whether or not the detection is correct.
Automation vendors account for this in their own guidance. Siemens, for example, notes that application or system instability in SIMATIC PCS 7 and PCS neo environments can result not only from malware, but from antivirus software’s reaction to a detection, whether it’s a true or false positive. Its current guidance also distinguishes recording a detection from automatically remediating it through deletion or quarantine.
Endpoint tools built for OT start from different design choices. They’re built to understand industrial protocols and the normal behavior of the processes around them, which helps them avoid the mistakes IT agents designed around enterprise environments can make with legitimate safety protocols or control-system commands. They can run lightly, mostly in user space, so the host’s critical tasks keep the resources they need. Many also support the older operating systems OT still runs and keep working on hosts with limited or no connectivity. What they act on can be limited, for example to new or changed executables that lack a trusted signature, and teams can decide whether a detection triggers an action or only an alert. Some are even being developed with automation vendors themselves.
Consequence also runs in both directions. A response can disrupt safety and production, and so can a threat left alone. So it’s worth asking the people who run the process: what happens if this asset slows down, restarts, or drops off the network? Their answers are what let a team set responses it can trust. Building that trust also takes transparency. Teams should be able to see how each response would behave on their actual assets, dependencies, and operating conditions before it goes live.
3. What Context Should Arrive with an Alert?
Picture an alert landing mid-shift, with production running. The person approving a response has to decide fast, so everything they need should come with the alert. That starts with knowing what happened on the host, how far it reached on the network, and what the asset is doing for the process right now.
The host tells you what happened. Remember the USB device from the first question? Nozomi Networks Labs research on BadUSB keystroke injection found that these attacks need no network connection at all, so network monitoring alone can miss them. On the host, though, the attack looks like someone typing faster than any person could. That can look different from normal operator activity.
The network tells you how far it went. Did the activity stay on that one machine? Did it reach other assets or touch the communications the process depends on?
The process tells you what is at stake. What is the asset doing at this moment? Is there a standby system or a manual fallback? The people who run the plant know these answers better than anyone, so it’s worth deciding before deployment how that knowledge reaches whoever approves the response. With all three, the approver can see what happened, how far it went and what acting on it could mean for the process. The most useful alert brings the host and network views together, so no one has to piece them together from two tools.
4. Which Actions Should the Tool Take on Its Own, and Which Need Approval?
Think about the asset classes in your environment and the actions an endpoint tool could take on each. Which would you be comfortable letting it take on its own? Which would you want a person to approve first? And where would an alert be enough? For most teams, the answers depend on how critical the asset is and how each action could affect the process.
SANS guidance on ICS incident response draws a clear line between IT and OT response. An IT playbook can treat isolating a host or killing a process as a routine first move. In OT, aggressive containment or an automated shutdown can halt production, damage equipment, or create unsafe conditions.
How sure you are about a detection and how much the tool should be allowed to do about it are separate decisions. A high-confidence detection makes a strong case for acting quickly. Whether the tool should act without a person often comes down to what that action could mean for the process and whether a standby system or manual fallback is there if it doesn’t go as planned.
For actions that need a person’s sign-off, it’s worth thinking through who that person is and how you’d reach them mid-shift. Do they know the process well enough to decide quickly? At a remote or unstaffed site, would they be reachable in time?
An alert isn’t the lesser option. For some assets, knowing right away that something changed is exactly what you need, and giving a tool’s policy free rein to quarantine or isolate isn’t worth the risk. Whatever your answers, those calls belong to your team, not to a product’s defaults, and an endpoint tool built for OT should let you set them by asset class.
5. How Will You Manage the Tool’s Updates Over Time?
Updates are how a tool keeps up with new threats. Each one can also change how it behaves, so it’s worth thinking about how those updates will reach your operational assets. It’s a familiar question in OT, where many assets run for decades and changes to them are planned carefully rather than rolled out as they arrive. The same planning should apply to security tool updates.
Vendors release updates on their own schedule, and plants make changes on theirs, usually around planned maintenance windows. The July 2024 incident started with a content update that was pushed out automatically. Do you hold updates for the next maintenance window and accept that detection content may fall behind? Or apply them as they arrive and find another way to fit them into your change process? The right balance can differ from one asset class to the next.
Whatever tool you choose, its updates should fit your change process, not the other way around. In regulated environments, that carries compliance weight too. NERC CIP, which applies to North American electric utilities, treats intentionally installed software as part of a system’s baseline configuration and requires changes to that baseline to be authorized and documented.
When testing, approval and rollout are settled ahead of time, the people running the process know what to expect from the tool after every update.
Building on What OT Teams Already Rely On
Your answers will look different for each asset class, criticality level and site, and they’re worth revisiting as an asset’s role changes. Working through them before deployment gives security and operations a shared basis for deciding where endpoint security belongs, how it should behave, and how its response is governed.
OT teams don’t have to choose between seeing what happens on their hosts and keeping the process running. Where it adds value, endpoint security built for OT can do both. That means control over what the tool does on its own, enough context arriving with the alert and updates that fit the change process operations already run, all building on the network monitoring OT teams already rely on.
Gartner Research on Endpoint Protection in CPS
Get complimentary access to the research →
Gartner, Restrict IT-Centric Endpoint Protection Tools in CPS Environments, Deepak Mishra, Sumit Rajput, Evgeny Mirolyubov, Katell Thielemann, 21 September 2026.
Gartner is a trademark of Gartner, Inc. and/or its affiliates.







