Most utilities file DDoS under IT.
The control systems sit behind segmentation, the plant runs whether or not the website loads, and the threat model stops at the corporate firewall. On paper, a flood is a reputational nuisance.
The two incidents that have actually touched grid operations through denial of service read differently. In neither did the attack need to reach a controller. One took away the operators' view of their own generation sites. The other stopped customers telling the utility that their lights were out.
Scope: the IT edge of a physical operation
DDoS testing for energy and utilities is an authorised exercise that measures how an electricity, gas or water operator's internet-facing systems and its IT/OT boundary behave under denial-of-service load, without sending any test traffic into operational technology. It covers the customer channels that matter during a physical incident (outage reporting, prepayment top-up, contact centre), the boundary services that carry plant data out to people (historian replicas, dashboards, remote engineering access), and the outbound paths that submit schedules to market and system operators.
The pass mark is not page uptime. It is how old the control room's picture became, how long an engineer waited to connect, and whether a customer in the dark could still report it or add credit.
Authorisation, vector choice and the reading of results do not change by sector, and the complete guide to DDoS testing covers them once. What follows is what changes when the system behind the website is a grid, a pipeline or a water network.
Where the test stops
Draw the line on the architecture, not in the contract
Most operators describe their network in Purdue terms: enterprise IT at Level 4, a DMZ for IT/OT data exchange at Level 3.5, supervisory systems at Level 3, controllers at Level 2, and field devices and the physical process below that.
A DDoS test ends at the IT-side face of the Level 3.5 DMZ. Nothing below it is ever a target, and nothing in the test plan should depend on a firewall rule holding to keep it that way. If the only thing standing between a test generator and a controller is a policy you are also testing, the scope is wrong.
What each zone gets
| Zone | Targeted? | What the test does there |
|---|---|---|
| Internet edge and customer channels | Yes | floods and application-layer load at agreed ceilings |
| Enterprise IT (Level 4) | Yes | load on remote access, SSO and dashboards as users see them |
| IT/OT DMZ, IT-facing side (Level 3.5) | Yes, carefully | historian replicas, jump hosts, file transfer, patch servers |
| Supervisory and control (Levels 3 to 0) | Never | watched by operations for latency, alarms and data flow |
The last row is not a formality. Industrial protocols such as Modbus/TCP and DNP3 were designed for closed networks, and many field devices handle unexpected traffic badly: a burst that a web server would shrug off can stall a controller's network stack. That is a reason to keep the test out, not a finding to go looking for.
Watching the OT side while you test
The OT side still belongs in the run, as an observer. Operations staff watch supervisory-to-controller latency, alarm rates and historian data flow for the whole window, and any OT alarm the team cannot immediately explain stops the test.
That is how the IT/OT segmentation gets evidence rather than an assumption: the flood ran at the edge, and the plant side showed nothing.
The operators' view is the first casualty
The plant keeps running when the picture freezes
Generation, transmission and treatment processes are controlled locally. A protection relay does not wait for the corporate network, and a pumping station keeps to its setpoints when the link to head office is down.
What the IT side carries is the human view: historian replicas that engineers query, operational dashboards, alarm summaries forwarded to on-call staff, and reports to the system operator. A flood that takes those away does not stop the process. It leaves people running it with an older picture than they think they have.
The incident that proved it
On 5 March 2019, a renewable generation operator in the western United States, later reported to be sPower, lost communications between its control centre and several generation sites in repeated short outages over a period of hours. The cause was a denial-of-service condition: an unpatched vulnerability in the web interface of its perimeter firewalls let an outside party reboot them over and over.
Generation itself was not interrupted. NERC's published lesson from the event was about firewall firmware and exposed management interfaces, and it is the first widely reported case of a DoS disrupting grid operations in the US. What the operator lost was sight of its sites, which is the outcome a utility test should be built to measure.
One device facing both ways
The sPower firewalls carried internet-facing traffic and the control centre's link to its sites. Smaller utilities, municipal operators and water companies often have the same arrangement: one firewall pair, or one internet circuit, in front of both the corporate edge and the remote-site telemetry.
Any stateful firewall on that path has a connection table and a CPU that a flood can fill. When the same box carries plant communications, an edge flood becomes an OT availability event without a single packet crossing into the OT network.
Data diodes protect one direction
Unidirectional gateways are increasingly common at the IT/OT boundary: data leaves the plant through hardware that cannot carry anything back. They are a strong control, and a flood on the IT side cannot travel through one into the plant.
They do not protect the receiving side. The historian replica and every dashboard reading from it sit on IT infrastructure, and they are the part the people use.
Measure data age, not page load
The useful metric for this layer is data age: the difference between the timestamp on the newest value an engineer can see and the time on the wall. A dashboard that loads in two seconds and shows readings twelve minutes old has failed, and a page-load check will mark it as healthy.
In the test window
Load the IT-facing historian API and the dashboard tier from outside, at the agreed ceiling, and record data age on the dashboards every thirty seconds. Separately, put load on the internet side of any firewall or circuit that also carries site telemetry, and have operations record telemetry gaps at the control centre. The second test is the one most utilities have never run.
Remote access is how engineers reach a site at 3 a.m.
One door, by design
Remote access into an electricity operator's control environment is deliberately narrow. NERC CIP-005 requires interactive remote access to high and medium impact systems to pass through an intermediate system with multi-factor authentication, and most operators elsewhere follow the same pattern: a jump host or remote access gateway in the DMZ, and nothing else.
That narrowness is the point for security. For availability, it means a single service in front of every remote engineer, and it faces the internet.
Slow attacks fit the choke point
The VPN layer's own failure modes (handshake floods, the authentication backend's ceiling, the failover storm) are set out in VPN DDoS testing. The utility-specific risk sits one hop further in, at the jump host.
Jump hosts are often configured once at commissioning and left alone. Long idle timeouts and small connection limits, chosen for a handful of engineers, are exactly what a Slowloris-style attack exploits: a few thousand slow connections at negligible bandwidth hold every slot, and the next engineer cannot log in.
Vendor access shares the same door
Turbine manufacturers, inverter suppliers and control-system vendors often support their equipment remotely, sometimes for whole fleets. Their sessions arrive through the same gateway, or through a vendor access platform with its own internet exposure.
When the KA-SAT satellite network was disabled by a wiper attack on 24 February 2022, a German operator lost remote monitoring of about 5,800 wind turbines, which kept generating. That was not a DDoS. It shows how much remote operation now rides on a link that someone else runs.
In the test window
Measure the time from "an engineer needs to connect" to "an engineer is on the jump host" while the gateway is under load, for an employee and for a vendor account. Then check the out-of-band path: a separate circuit or a cellular route that bypasses the attacked edge, and when it was last actually used.
Customers call during the outage, not after it
Ukraine, December 2015
On 23 December 2015, attackers opened breakers at substations belonging to three Ukrainian distribution companies, and about 225,000 customers lost power for several hours. The joint analysis published afterwards records that, at the same time, the companies' call centres were flooded with calls, a telephony denial of service that kept real customers from getting through to report the outage.
The intrusion into the control systems was the attack. The call flood was there to slow the response, and it needed nothing more sophisticated than volume against the channel customers use first.
A storm is the legitimate version
The utility's biggest legitimate surge has the same shape. A storm takes out supply to a few hundred thousand homes, and those households go straight to the outage map, the app and the phone line, on mobile data, from phones they are trying to keep charged.
Outage maps are often live: every view queries the outage management system for current incidents, polygons and restoration estimates. A map built that way multiplies storm demand straight into the database.
The arithmetic of a storm night
Take an illustrative model. On a calm evening the map draws about one view every five seconds, and each view makes twenty API calls for incidents, polygons and estimates: about 4 requests a second at the origin.
Now let a storm cut supply to 300,000 homes. If 40% of them check the map and refresh it every ten minutes, that is 200 views a second, or 4,000 requests a second at the origin. The same map published as a set of static files every 60 seconds and served through a CDN shield costs the origin about one request every three seconds, however many people are looking.
The general case for serving from cache under load is made in CDN cache absorption. For a utility the point is narrower: restoration estimates change every few minutes at most, so nobody needs them computed per view.
The outage footprint tells you who is real
Detection tuned on calm evenings will flag a storm surge. Utilities hold something most sectors do not: the outage management system knows which postcodes and feeders are off supply right now.
Legitimate storm traffic comes from those areas, largely over mobile networks, and asks about those incidents. A flood does not know the footprint. Feeding the current outage area into the edge rules gives a discriminator no attacker can copy in advance.
The phone line is the old channel and the first target
Contact centres now often run on SIP trunks over the same internet circuits as everything else, which puts the 2015 technique within reach of an ordinary SIP flood as well as a call flood. Ask whether the outage line has an announcement that answers without an agent, and whether emergency numbers (gas escapes in particular) ride a separate path from the general line.
Prepayment customers run out of credit, not patience
The meter keeps counting
A prepayment customer whose top-up fails does not get an error page and come back later. The meter keeps drawing down credit. Emergency credit buys some time, many meters have friendly-credit hours in which supply is not cut, and after that the household is off supply until a top-up lands.
That makes the vend path a flood target with a physical consequence, and the households on it include many who can least afford a lost evening of heat.
The vend path has more hops than the payment
A top-up runs from the app or website, through payment authorisation, to the supplier's vend system, through the meter head-end, and across the communications network to the meter. In Great Britain that last leg is a shared national smart-meter network, which no supplier may load in a test.
The fallback matters as much as the main path. Smart prepayment meters can usually accept a top-up code typed on the meter's keypad when the remote credit message does not arrive, and older key and card meters top up at a shop. Both bypass the attacked web tier.
In the test window
Flood the web and app tier and measure completed top-ups per minute end to end, not payments accepted. Check that payments taken during the attack either credit the meter or are reversed, and that the keypad-code fallback is issued and works. The supplier's own vend queue is in scope; the shared network beyond the head-end is not.
Market submissions leave through the same circuit
Gate closure does not move
Generators, suppliers and traders send schedules, bids and notifications to market and system operators against fixed gate closures. In Great Britain the Balancing Mechanism closes one hour before each half-hour settlement period; the European day-ahead coupling closes at noon Central European Time on the day before delivery.
A submission that misses its gate leaves the firm exposed to imbalance prices for whatever it failed to schedule. Banks face a similar problem with payment cut-offs, and DDoS testing for financial services handles that side. The energy version has a different weak point.
The submission is outbound
Market submissions travel out from the firm, over the same internet circuits and through the same firewalls an inbound flood fills. A saturated link drops the outbound traffic too, and an edge filter that blocks or challenges unfamiliar flows can catch the API responses coming back.
Test from the trading desk's real systems: under load at the edge, does a submission still reach the operator's interface and get its acknowledgement back, and does the documented backup route (a second site, a second circuit, the operator's own fallback channel) actually work when used?
Water utilities: smaller edges, the same boundary
Water and wastewater operators have fewer online transactions than energy suppliers, and their public site matters most on the day a boil-water notice or a burst main sends everyone to it at once.
Their exposure sits at the boundary. Remote pumping stations and treatment sites often report over cellular or private radio to a control centre that shares an internet edge with corporate IT, and remote access for contractors is common. The test boundary, the data-age measurement and the out-of-band check above apply unchanged; the outage-map arithmetic usually does not.
What NERC CIP, NIS and NIS2 have to say
None of these frameworks names DDoS testing. They do make availability of essential services and the security of the boundary a regulated matter.
In North America, the NERC CIP standards are mandatory for the bulk electric system: CIP-005 governs electronic security perimeters and remote access, and CIP-008 and CIP-009 cover incident response and recovery plans that are expected to be exercised. In the UK, energy and drinking water operators designated as operators of essential services fall under the NIS Regulations 2018 and are usually assessed against the NCSC's Cyber Assessment Framework. Across the EU, Directive (EU) 2022/2555 (NIS2) puts electricity, gas, oil, hydrogen, district heating, drinking water and waste water in its Annex I, which brings larger operators under its risk-management and incident-reporting duties. US community water systems also have risk and resilience assessment duties under America's Water Infrastructure Act of 2018.
This is not legal advice. A test record that shows the boundary held, how old the operators' data got, and that a runbook was followed is usable evidence under each of them.
A run sheet for a utility test
Traffic ceilings and the stop button are generic; testing without disrupting production has them. The steps below are the ones a grid or water operator has to add:
- Agree the boundary in writing with the OT owner, by asset and address, not by network name. Anything below Level 3.5 is listed as excluded.
- Book the window against the weather and the grid. Avoid forecast storms, heatwaves and cold snaps, and any period when the system operator has issued a capacity warning. Have a named person who can call it off for a real event.
- Put operations in the room, watching OT latency, alarms and telemetry gaps, with authority to stop the test on any unexplained alarm.
- Tell the upstream provider and the market-facing team. A trading desk that sees submission latency climb mid-test should know why.
- Run each surface separately first: customer channels, historian and dashboards, remote access, then the shared edge. Combine them only once each has a known ceiling.
- Measure the four numbers that matter here: dashboard data age, time to a connected engineer, completed top-ups per minute, and outage reports accepted per minute.
- Exercise the runbook, not only the network. Contact lists, vendor escalation and who approves a customer message go stale faster than firewall rules; a DDoS tabletop exercise before the live window finds most of it.
- Check what drained afterwards: queued top-ups credited, historian backfill complete, outage reports from the window present in the outage management system.
FAQ
Where does a DDoS test stop in an OT environment?
At the IT-facing side of the IT/OT DMZ. Supervisory systems, controllers and field devices are never targeted, because industrial protocols and devices can fail on traffic a web server would ignore. They are watched during the test, and any unexplained alarm stops it.
Does a data diode or unidirectional gateway protect against DDoS?
It stops a flood travelling from IT into the plant, which is valuable. It does nothing for the historian replica, dashboards and alerting on the IT side, which is where engineers read the plant's state. Those still need testing, measured by how old the data they show becomes.
What did the 2015 Ukraine grid attack have to do with DDoS?
The outage itself came from attackers opening breakers through compromised control systems. Alongside it, the distribution companies' call centres were flooded with calls so customers could not report the outage. It is the clearest public case of a denial-of-service used to slow a utility's response rather than to cause the incident.
Why do outage maps fail during storms?
Many query the outage management system on every view, so demand at the origin scales with the number of people off supply and how often they refresh. Publishing the map as static files on a short interval and serving them from a CDN removes that link between storm size and origin load.
Which regulations cover DDoS resilience for energy and water utilities?
NERC CIP in North America for the bulk electric system, the NIS Regulations 2018 in the UK, and NIS2 in the EU, plus America's Water Infrastructure Act for US community water systems. None names DDoS testing; all expect essential services to stay available and their controls to be exercised.
The meter you cannot read
Electricity and water operators spend a great deal on knowing the state of their network to the second. Under a flood, that knowledge depends on a chain of IT systems most of them have never loaded: the replica, the dashboard, the gateway, the circuit the telemetry shares with the website.
A grid can ride through a DDoS without a single breaker moving. The open question is whether, while it happens, the people responsible for it can still see it, reach it and hear from the customers on it.
That can be measured on a quiet afternoon. The alternative is learning it on a storm night, from the call centre.
