Explore our new Cybersecurity Terms Glossary, explained by our experts.

View the Glossary View the Glossary
PRIVIA

Cybersecurity Terms Glossary

What Is Anomaly Detection in OT?

Anomaly detection in operational technology identifies deviations from a learned baseline of industrial network traffic and process behaviour, and it works far better here than in IT because the baseline stays stable.

Illustration for the cybersecurity term OT Anomaly Detection
Home / Blog / Cybersecurity Terms / What Is Anomaly Detection in OT?
17 November 2026

What Is Anomaly Detection in OT?

What Is Anomaly Detection in OT?

Anomaly detection has a mixed reputation in enterprise security, and deservedly so: corporate networks change constantly, users behave unpredictably, and a baseline of normal is obsolete by the time it is built. Industrial networks are the opposite. A plant network is a small, fixed population of devices exchanging the same messages in the same order at the same intervals, often for years. Learning what normal looks like is realistic, and departures from it are meaningful.

That property is why passive monitoring is the single most effective detection investment in an operational environment, and why it works where agents and scanners cannot be used at all.

How It Is Deployed

The sensor is passive by necessity. Traffic is copied from a span port on a switch or, preferably, from a network tap, so the monitoring equipment can fail without affecting the process and generates no traffic of its own. Nothing is installed on controllers, operator stations, or safety systems.

The sensor parses industrial protocols rather than merely counting packets: Modbus, DNP3, IEC 60870-5-104, IEC 61850 including GOOSE, EtherNet/IP, PROFINET, S7comm, BACnet, and OPC UA. That protocol awareness is what separates a useful OT sensor from a network monitoring tool, because the interesting events are expressed in protocol semantics rather than in volumes.

What It Actually Detects

  1. New assets and new conversations. A device that has never been seen, or a pair of devices that have never spoken, in a network where the device list has not changed since commissioning.
  2. Engineering operations. Program downloads to a controller, firmware updates, configuration changes, and a controller switching from run to program mode. These are rare, high-consequence, and entirely visible on the wire.
  3. Unexpected commands. A write function code from a source that only ever reads, a control command on a conduit designated read-only, or a command issued outside the normal operating pattern.
  4. Protocol-level manipulation. Spoofed GOOSE publishers with impossible state numbers, duplicated identities, and malformed messages consistent with fuzzing.
  5. Process-aware anomalies. The most advanced layer: a setpoint outside the operating envelope, a sensor reading physically inconsistent with its neighbours, or a valve position that does not match the flow it should produce. This catches manipulation that is perfectly legitimate at the protocol level.
  6. Known indicators. Signature and rule matching against published industrial threat content, which covers the specific tooling described under Industroyer and Triton.

The tooling ranges from commercial industrial detection platforms through open-source network security monitors with industrial protocol parsers to the free network monitoring suite published by a US national laboratory, which is a credible starting point for a site with no budget.

The Pitfalls

Three recur in every deployment. Baselines learned during commissioning, a shutdown, or a turnaround are not baselines of normal operation, and a sensor trained in the wrong week will alert continuously afterwards. Span port capacity is routinely underestimated, and a sensor silently dropping traffic produces confident, incomplete conclusions. And serial segments are invisible unless taps are fitted, which means the field level beneath a gateway may be entirely unmonitored while the dashboard shows green.

There is a fourth, less technical. Detection without a response plan is a dashboard. If an alert at 03:00 reaches nobody with the authority to act on it, the investment has bought evidence for the post-incident report and nothing else. The alerts need to land in the same place as everything else, with analysts who have enough process context to tell a planned engineering change from an intrusion, and an engineering escalation contact who expects the call. That is the collaboration problem addressed under OT and IT working together.

Making It Pay for Itself

The first benefit is usually not detection at all. A passive sensor produces the asset inventory that the site has never had: every device, its vendor, model, firmware version, and communication relationships, built without sending a single packet. That inventory then feeds the zone and conduit design under IEC 62443, the vulnerability management programme, and the evidence a regulator asks for.

Useful metrics follow: unexpected engineering operations per month, unauthorised assets found, conduit rule violations, and mean time to triage an industrial alert. For UK operators of essential services, demonstrable detection capability in the operational environment is an explicit expectation of the NCSC Cyber Assessment Framework, so the output has a regulatory audience as well as an engineering one.

A closing observation worth keeping in mind. Triton was discovered because the safety system tripped, not because anyone detected the intrusion that had been under way for months. The network traffic that would have revealed it existed the whole time.

At Privia Security, we design and tune passive detection in industrial environments through our SOC Maturation Services.

FAQ – Frequently Asked Questions

Why Does Baselining Work Better in OT Than in IT?

Because industrial traffic is genuinely repetitive. A plant network has a fixed set of devices exchanging the same protocol messages at the same intervals, often unchanged for years, so a learned baseline stays valid. Corporate networks change daily as users, applications, and cloud services come and go, which makes a stable definition of normal far harder to establish.

Does an OT Sensor Need an Agent on Controllers?

No, and that is the point. Detection is passive, taking a copy of traffic from a span port or network tap, so nothing is installed on controllers, operator stations, or safety systems and nothing is transmitted onto the network. This is what makes it usable where vendor support agreements and safety certification prohibit third-party software.

What Is Process-Aware Anomaly Detection?

Detection based on the physical plausibility of process values rather than on network behaviour: a setpoint outside the safe operating envelope, a sensor reading inconsistent with neighbouring measurements, or a valve position that does not match the flow it should produce. It catches manipulation that is entirely legitimate at the protocol level, which is precisely what a capable attacker would use.

When Should the Baseline Be Learned?

During a representative period of normal production, not during commissioning, a shutdown, a turnaround, or a seasonal extreme. A baseline learned in the wrong week produces continuous false alerts afterwards and trains the team to ignore the system. Plan for a re-baseline after any major plant modification, and record what the baseline period covered.

Related Terms

You May Be Interested In These

What Is tfsec?

Cybersecurity Terms

17 November 2026

What Is tfsec?

Read More
What Is Red Team?

Cybersecurity Terms

30 March 2021

What Is Red Team?

Read More