Nginx on DNP3 Port 20000, Again: Why the Nov 19 Score-7 Hits Were CDN Edges, Not Grid Gear
Our anomaly hunter scored two US hosts a 7. The banners explain why, and it is not pressure data.
Jeff Gray · · Scan of 19 November 2026 · Control Systems Security
Illustrative image, not a photo of any scanned host.
Our twice-weekly Global ICS Anomaly Hunter ran on 19 November 2026 at 23:00 UTC. Its two highest scores were a pair of 7s: An nginx web server answering on TCP port 20000, the standard DNP3 port, once in Los Angeles and once in Pullman, Washington.
On paper that looks bad. DNP3 is the protocol utilities use to talk to substations, pump stations, and remote terminal units. A web server on that port, plus a scorer note that said "Exposed control metric: PSI", reads like pressure telemetry leaking onto the internet.
It is not. When we pulled the raw banners, both hosts turned out to be content delivery network edge nodes, and the "PSI" came from a random request ID. This post walks through how we got to that answer, because the same check will save you time on your own exposure findings.
What the scan saw
The run used four Shodan index queries and processed 63 host records:
| Query block | Records | Flagged | Top score |
|---|---|---|---|
| DNP3, TCP 20000, United States | 20 | 5 | 7 |
| Modbus, TCP 502, China | 20 | 20 | 4 |
| Siemens S7, TCP 102, Iran | 16 | 3 | 4 |
| IEC 60870-5-104, TCP 2404, Russia | 7 | 1 | 4 |
That is 29 newly flagged hosts, with 17 more suppressed because the same address and port had already been seen in an earlier scan. Every one of the 20 Modbus hits sat on a large commercial cloud provider, which is why they scored 4: Cloud-hosted Modbus is usually a lab, a honeypot, or a test rig, not a plant.
Why the two DNP3 hosts scored 7
The hunter adds points for things that look out of place. These two hosts picked up:
- +3 for a consumer or IT service (nginx) on an industrial port.
- +4 for an "exposed control metric," triggered by the string
psiin the banner.
Both hosts sit in the same autonomous system, AS54994, and both returned nearly identical responses.
What the banner actually says
Here is the Los Angeles response with the request ID lightly masked. The Pullman host returned the same shape:
HTTP/1.1 400 Bad Request Server: nginx Content-Type: text/html Content-Length: 2360 Connection: close x-ws-request-id: 6aff7ee3_PSxjpSin5ah152_37654-2429
Three things stand out:
- It is plain HTTP. A real DNP3 outstation does not answer with an HTTP status line. DNP3 frames start with the two bytes
0x05 0x64. Nothing here speaks DNP3. - The
x-ws-request-idheader is a CDN fingerprint. A large commercial content delivery network stamps it on every response, and the middle segment names the edge node that served the request. AS54994 is part of that CDN's footprint, and its edge nodes answer HTTP on many ports, 20000 included. - The "PSI" is inside the node ID. The segment
PSxjpSin5ah152contains the lettersp-s-iin a row. Our scorer lowercased the banner and searched forpsias a substring, so a random identifier became "pressure in pounds per square inch."
Put together, the honest reading is simple: These are CDN edge servers that happen to listen on port 20000, and the 7 is a scoring artifact. There is no evidence of a utility, a controller, or any process value.
The pattern behind the pattern
This is not the first time our hunter has flagged nginx on port 20000 from this address space. It has done so in most scans since August, under a few different organization labels, with the same "PSI" reason attached to the 7s. Earlier write-ups of this pattern described the PSI match as exposed telemetry. The banner evidence above points to the same artifact, and we are reviewing those posts.
That is the real lesson of this scan. A recurring "anomaly" that never changes shape is usually a property of your detector or of shared infrastructure, not a growing threat.
What to do with findings like this
If you run your own exposure monitoring, or you are reading someone else's, these checks separate real industrial exposure from noise:
- Verify the protocol, not the port: Before calling something DNP3, confirm it answers a DNP3 link-layer request with a frame that begins
0x05 0x64. An HTTP status line on port 20000 is a web server, full stop. - Check who owns the address first: CDN and cloud ranges that answer on ICS ports are almost always generic listeners. Look up the autonomous system before you write the headline.
- Match process values as words, in context: "PSI," "MW," "Hz," and "kWh" should only count when they appear as labeled values in a page body or protocol field, never as substrings of IDs, hashes, or cookies.
- Treat a repeating hit as a detector question: If the same finding comes back every scan with the same reason, test the rule before you escalate the host.
And for your own estate, the boring advice still holds: Real DNP3 outstations should not be reachable from the internet at all. Put them behind a managed, authenticated remote access path and alert on any inbound session to TCP 20000 from outside your control network.
Method: This analysis uses passive Shodan index records only, collected by our automated Global ICS Anomaly Hunter on 19 November 2026. No active scanning, probing, or connection attempts were made against any host. IP addresses are withheld and organization names are limited to public network ownership records. Scores come from an automated rule set and are a triage signal, not a finding on their own.