How it works
How Northwatch is being built to defend a site.
This page describes how Northwatch is being built to work. Most of it is a plan, not a finished product. Every part is labeled, so you can see what exists in the lab today and what is still a plan.
- In the lab today
- Built and running in our lab now. Today that is one detector, trained on a public recording of a test plant, and the independent tester that scores it.
- Partly in the lab
- A first piece exists in the lab. The rest is planned.
- Planned
- A design goal. It has not been built or tested. Nothing on this page has been tested at a real site.
Four steps
What Northwatch is designed to do, in order.
Network protection comes first. Machinery comes second.
-
Learn what normal looks like
It starts by listening to your site.
The system plugs into your network and reads a copy of the traffic, the messages your devices send each other. It sends no commands to your machines.
From that it is designed to build a map: which devices exist, what talks to what, and how each line usually behaves on each shift. That includes the controllers, the small computers that run your lines.
Partly in the lab The lab detector learns what normal sensor readings look like on a recorded test plant. Mapping a site is planned.
-
Notice what does not belong
Anything that breaks the pattern of a normal day.
Once it knows normal, it can tell when something is off: a device that speaks to a controller for the first time, a command sent at an odd hour, a pump whose readings drift in a way they never have.
It is being built to judge an action by what it would do, not by who appears to have sent it.
Partly in the lab A detector that watches sensor readings and raises an alarm when the plant behaves abnormally is being trained now. A network traffic detector is next.
-
Stop it with the defenses you already own
Northwatch spots the attack. What you already own blocks it.
You most likely own a firewall, the gatekeeper that decides which network traffic is allowed through, and protection software on your computers. Northwatch is designed to tell them what to stop.
It does not send commands to your machines. Any action with physical consequences waits for a person to approve it.
Planned Nothing in the lab blocks anything yet. A pilot system would only listen.
-
Alert your team in plain words
One clear alert that a shift supervisor can act on.
Every alert is designed to say four things: what happened, what it would have done physically, what Northwatch did, and what you must do now.
No security training should be needed to read it. Many small events that belong to one attack should arrive as one alert, not thousands.
Planned The alert design exists on paper. See an example on the product page.
In the lab today
One detector, measured by an independent tester.
What exists now is the first piece of step two: a detector that watches the sensor readings of a recorded test plant. It is trained and scored on public data. A separate testing program does the measuring.
of attacks caught, with false alarms per day. A false alarm is an alarm when nothing was wrong.
These are lab numbers on one public recording of a test plant (, license , ). They do not prove that Northwatch is ready for a real site.
The lab numbers could not be loaded on this page. Please try again, or open the lab results page.
Design goals
What Northwatch is being built to do.
These are goals, grouped into four themes. They are not a list of shipped features. The label beside each one says where it stands today.
Know the site
You cannot guard what you have not mapped.
-
Map every device and connection
Build a living map of every device, controller, account, program and connection on the site, and keep it up to date as the site changes.
Planned
-
Rank the weakest links
Find every way in and every spot where one failure would spread, then rank them. Northwatch assumes attackers find the same list.
Planned
-
Know what a command does physically
Keep a simulated copy of the site, a model of the plant built in software. A risky command is run through the copy first, before it reaches a machine.
Planned
-
Keep a verified clean copy of every device
Hostile code can sit silent for months, so watching behavior is not enough. Each device is compared against a copy known to be clean, and any difference is flagged.
Planned
See the attack
From a flood of attempts to one quiet command.
-
Connect every layer
Network traffic, sensor readings, logs (the records devices keep of what they did) and code are read side by side, because an attack often looks normal in any one of them. Today the lab detector reads sensor readings only.
Partly in the lab
-
Adapt from the smallest attack to the largest
Catch mass attacks that arrive in waves, and also the single quiet probe, a small test for a weak point, that looks like nothing on its own. Every attack that works in the lab is meant to become a permanent test, so the same trick does not work twice.
Planned
-
Inspect new code before it runs
New code is read, then run in a sealed test box, a closed space where it can do no harm, while Northwatch watches what it really does.
Planned
-
Hold up against swarms of AI agents
AI agents are programs driven by AI that act on their own. A swarm is many of them working at once. Northwatch is being built to link many small probes into one attack and respond to the whole. One swarm should produce one clear alert, not thousands. It is also being built to limit how fast anything can try anything, so the speed of a machine stops helping the attacker.
Planned
-
Double-check your other defenses
Northwatch is being built to work alongside every security system you already run. If one of them misses an attack, or a system goes down without warning, Northwatch is a second and third check that can still notice it.
Planned
Respond safely
A wrong response can hurt people too.
-
Fixed safety rules first
Simple rules that do not wait for an AI model react in milliseconds. AI reasoning sits above them and cannot overrule them.
Planned
-
Block through the defenses you already own
Northwatch tells your firewall and device protection what to stop. They do the blocking, and nothing you paid for is thrown away.
Planned
-
Isolate and rebuild
A device that has been taken over is cut off and rebuilt from the clean copy. Cleaning it in place is not trusted.
Planned
-
Stop a cascade failure
A cascade failure is one failure that sets off the next, like a row of dominoes: a pump stops, the tank behind it overflows, and the line shuts down. Northwatch is being built to know which systems depend on which, and to contain the first failure before it reaches the rest.
Planned
-
Never cause the accident itself
Every response is checked against the safety rules and the simulated copy first. Any action with physical consequences needs a person's approval.
Planned
Stay honest
A good number that is wrong is worse than a bad number that is right.
-
An independent tester measures, not the builder
An automated loop tries one change to the detector at a time. A separate testing program measures it and keeps the change only if it is really better.
In the lab today
-
A hidden second exam
A second set of attacks that the loop never sees. It catches a detector that only memorized the first exam.
In the lab today
-
Every round is recorded
Each round is written down, including the ones that were thrown out, and published on the lab results page.
In the lab today
-
Five measures
Share of attacks caught, false alarms per day, seconds from attack to alarm, seconds from alarm to containment, and zero unsafe actions by Northwatch itself. The lab measures the first three today.
Partly in the lab
-
Alerts in plain words
Every alert is designed to say what happened, what it would have done physically, what Northwatch did, and what the operator must do now.
Planned
Brainpower in layers
Fast rules at the bottom. Deeper thinking only when it is needed.
An AI model is a program that has learned from examples. Large models reason well but are slow and costly to run. Safety cannot wait for one, so Northwatch is designed in three layers. All three are planned, and all three are designed to run on site, so your data stays in the building.
-
Layer 3
A large model, for hard cases only
Called when the layers below cannot decide. It reasons through the hard case, and what it concludes is designed to become new rules for the layers below.
- Used
- Rarely
- Speed
- Slowest of the three
-
Layer 2
A mid-sized model on site
Connects events across the whole facility. One odd login and one odd valve command may mean nothing apart and one attack together.
- Used
- All the time
- Speed
- Slower than rules
-
Layer 1
Fixed safety rules
Simple rules and very small models that react without waiting for anything above them. A rule here might be: this valve never opens while that tank is filling.
- Used
- On every command
- Speed
- Milliseconds, by design
The speeds above are design goals. None of these layers has been measured yet, so there are no numbers to show for them.
Safety limits
Security that causes an accident has failed.
Shutting down the wrong equipment can hurt people as surely as an attack can. These limits apply to everything Northwatch does.
-
01
It must never cause the accident itself.
Every response is designed to be checked against the fixed safety rules and the simulated copy of the site before anything happens. One of the five measures we score is zero unsafe actions by Northwatch.
-
02
A person approves physical actions.
Any action with physical consequences needs a human's approval, unless the site approved that exact action in advance.
-
03
It is defense, never offense.
War is not only offense. It is also defense, and defense is all Northwatch builds. All attack testing stays inside our lab, on recorded or simulated data. We do not probe real systems, and nothing in the lab sends commands to real equipment.
-
04
A real site needs human sign-off.
Before Northwatch runs at any real site, it needs an audit by a human security expert, and sign-off from the site owner and the site's safety engineers. A lab result does not replace either.
No security product stops every attack, and Northwatch will not either. The aim is to catch an attack in seconds, contain it, and prove how well that works with measured numbers.
Pilot sites
Help us take the next step out of the lab.
Northwatch is looking for a small number of mid-sized factories and distribution centers as pilot sites. A pilot system would only listen and learn. It would send no commands, and your data would stay in your building.