CVE-2026-58115: Node-RED's Open Editor on the IoT2050
Siemens rated a missing login on the SIMATIC IoT2050 Advanced at 10.0, and the default behind it ships with a low-code runtime plenty of plant gateways already run.
On 11 August 2026 Siemens ProductCERT published SSA-834709: one product, one missing login. CISA republished it two weeks later as ICSA-26-237-03. The device is the SIMATIC IoT2050 Advanced, the small gateway that sits in a control panel and moves process data to whatever lives upstream of it. On units running Industrial OS with Node-RED installed, below V4.3.4.1, the Node-RED HTTP interface doesn't enforce authentication. Anyone who can reach the port can build a flow and run commands on the device at maximum privilege. Siemens scored that 10.0, the top of both the CVSS v3.1 and v4.0 scales, and filed it under CWE-306, missing authentication for a critical function.
A score that high usually means several conditions lined up for the attacker. Not here. The flow editor on that gateway behaves the way its upstream project documents it behaving, and the Node-RED documentation is blunt about what that is: by default the editor is unsecured, and the project states that this "is only suitable if you are running on a trusted network." Siemens shipped a device where that default met a plant network. The part number matters to whoever orders the patch. For everyone else the number worth having is how many boxes in the plant run the same low-code runtime, installed by somebody who needed a Modbus register in a dashboard by Friday.
A wiring canvas with a shell behind it
Node-RED is a browser-based tool for wiring data paths. You drag nodes onto a canvas, join them with wires, press Deploy, and the runtime executes what you drew. Two node types in the stock palette decide how much a missing login costs. The exec node runs a shell command and hands back stdout. The function node runs JavaScript inside the runtime process. Neither is exotic. Both are how people do the small jobs a gateway exists for: rename a file before SFTP picks it up, prod a serial device that went deaf, reshape a payload so the historian accepts the tag name.
Deploying is an HTTP POST. The canvas in your browser is a client of the Admin API served on the same port, which means "unauthenticated editor" and "unauthenticated remote code execution" describe one condition in two vocabularies. An attacker never has to load the canvas. A single POST to /flows carrying an inject node wired to an exec node does the whole job, and on the Siemens image that command runs as root, because the service unit in the vendor's own Node-RED package specifies User=root with its working directory in /root.
The gateway has a foot in two networks
The CVSS vector marks a scope change, and the advisory says why in substance: a missing login on the Node-RED HTTP interface lets an attacker "execute arbitrary code on the underlying server with maximum privileges". On a panel gateway the reach doesn't stop at the server. A flow editor on a web server is a bad day for the web server. A flow editor on a panel gateway is different, because of where the gateway is bolted and what it's already connected to.
What these boxes do on a site is plain translation work. A gateway takes registers off a drive or a PLC, renames them into something a historian will accept, buffers when the link north drops, and forwards when it comes back. A common installation has two Ethernet ports in use: one facing the process segment with live S7, Modbus TCP or OPC-UA sessions, one facing the business network, the historian, or a router that someone added for remote support (often during a commissioning crunch, and rarely removed afterwards). Flows read and write. A Modbus write node pointed at a holding register can be a setpoint or a command word, and the runtime stores the credentials and certificates its flows are configured with, for every endpoint they talk to. So the prize behind an open editor on this class of device is rarely the data on the device. It's the authenticated path to equipment that's running: a burner control, a CIP skid, a conveyor drive, whatever the integrator wired up on the process side.
There's a second asset on the device that rarely makes it into a risk register. The flow holds the mapping: which Modbus register is the agitator's current, which S7 datablock offset is the batch number, which scaling factor turns a raw integer into degrees Celsius. On many sites that table is the only current written record of how the plant's addresses relate to the plant's vocabulary. Read access to the editor copies it.
Siemens had already changed this in its own package
Four and a half months before the advisory, on 25 March 2026, a Siemens engineer changed the Node-RED recipe in the company's open-source IOT2050 layer so that installing the package no longer enables or starts the service. The commit message gives the reason in one clause: Node-RED "currently runs with root privileges," so the maintainers decided users should start it deliberately. A one-line systemd preset file, disable node-red.service, carries the decision.
Read those two events in order and the triage question changes. Before that commit, installing the package brought up a root-privileged flow editor at boot, listening on every interface, with no login at all. After it, the same software sits on disk, inert, until an engineer runs systemctl enable. August's advisory then assigns a CVE and a version boundary to the earlier state.
Which gives you something to read alongside the version table. A unit below V4.3.4.1 with the package present and the service never started isn't listening today, but Siemens lists it as affected, because the condition in the advisory is "Node-RED installed", and one systemctl enable later it is the vulnerable device the advisory describes. Update it. A unit at V4.3.4.1 where the commissioning engineer installed Node-RED themselves, wrote a settings.js by hand and left adminAuth commented out is just as open, and no advisory covers that configuration, because an advisory can name a product while the default lives in the tool. The same low-code runtime is doing data plumbing right now on industrial PCs, on other vendors' gateways, on single-board computers taped inside a panel to log a flowmeter for a trial that no one formally ended. None of them appear in ICSA-26-237-03.
| What you find on the box | In the advisory's range | Exposed |
|---|---|---|
| Package not installed | No | No |
| Installed, service never enabled | Yes | Not listening until someone enables it; update anyway |
Service running, adminAuth commented out | Yes | Yes |
| Updated image, service running | No | Check the settings file the service loads |
| Installed from npm on any Linux box | No | Whatever the local settings say |
There's no second opinion to look up. CISA's version of this advisory is a direct conversion of the vendor's CSAF file, republished verbatim, with a note on the page disclaiming responsibility for its technical accuracy. No independent risk evaluation, no agency scoring. The 10.0 is Siemens assessing Siemens, which in this case is the right number, and it also means the triage work is yours from the start.
Looking for them in commissioning order
Start with the panel, not the asset register. The register lists what was purchased; the panel holds what was installed.
- Walk the doors. Note every box with an Ethernet port and a configuration web page: IoT2050s, industrial PCs, protocol converters, small Linux boards left behind after a trial, anything on a DIN rail that didn't come from the OEM's bill of materials.
- On each one, look for the process rather than a unit name:
ps -eo user,pid,args | grep -i '[n]ode-red', thensudo ss -ltnp. The Siemens image names its unitnode-red.service, the upstream install script names itnodered.service, and copies started by pm2, Docker or by hand have no node-red unit of their own. The default editor port is 1880; a different one comes fromuiPortin the settings file, thePORTenvironment variable or a--portflag on the command line, andssshows what is bound. - Look for listeners the walk-down missed, passively first: the switch MAC and ARP tables for the process VLAN, a SPAN port capture, the OT monitoring sensor if the site has one. If you probe actively, get operations' sign-off or a maintenance window, use a plain TCP connect check against the known editor ports at a low rate, and skip UDP, version and OS detection. Older PLCs, drives and serial converters have been hung by a broad scan, which is the outage you're trying to prevent.
- Open the settings file the running service actually uses; the runtime prints its path as
Settings filein the startup log (/root/.node-red/settings.json the Siemens image, whose service runs as root with its user directory under/root;~/.node-red/settings.jsby default elsewhere). Look foradminAuth. Commented out means no login, whatever the device's version number says. - Read the deployed flows for
execandfunctionnodes, and for write-capable Modbus, S7 and OPC-UA nodes. That pair tells you what an intruder would have inherited. - Ask the historian. Tags arriving from a source nobody can place are often a flow somebody built on a box that's now unmaintained.
Asking the historian finds boxes a walk-down misses. The forgotten instance is usually a success story: a flow that did its job, kept doing it, and outlived the person who wrote it.
What a deploy leaves on disk
A flow isn't a session. Pressing Deploy saves the flows to the flow file in the user directory, and the runtime loads that file again at every start, so anything an intruder installs survives a reboot, a power cut and a controller restart. An inject node set to repeat every few minutes is a scheduler that needs no cron entry. And outbound traffic from a gateway is the traffic a firewall rule already permits, because sending data north is the device's whole purpose, so a flow that calls out to an address no one recognises will not look obviously wrong on a flow log.
That makes a handful of forensic checks worth running on any box you find open:
- The modification time on the flow file, against the date of the last change anyone can account for. The runtime keeps a backup copy of the previous flow file beside it, as a dotfile a plain
lswon't show, and its timestamp is the second data point. - The service journal. Each deploy stops and starts the flows, and those events are logged, so a flow restart on a date with no work order needs an explanation. The journal doesn't say who deployed or from where. The runtime's audit log records the client address, and the user once
adminAuthis on, but it ships switched off, so a box you find open has none to read; turn onlogging.console.audit: truenow, for whatever comes next. - The node palette against the installed module list. A flow referring to a contributed node that nobody remembers adding is either an old experiment or somebody else's.
- Northbound byte counts per gateway over a few weeks. A device whose job is a dozen tags at one-second resolution has a predictable shape, and departures from it show up in the historian's own ingest statistics.
None of this proves a negative, and all of it looks only at Node-RED. The threat here is code running as root, and root doesn't need a flow to stay on the box: an SSH key in authorized_keys, a cron entry, a systemd unit, a changed binary, then the original flows put back and the timestamps with them. So if the editor was reachable from anywhere you don't trust, treat the host as compromised at the service account's privilege, which on the Siemens image is root, and assume escalation elsewhere until you can rule it out. Check the operating system's persistence points as well as the flow file. Rotate every credential and certificate stored in the flow credentials file, and every one the gateway uses towards the PLCs, the historian and the broker. Prefer reimaging to cleaning up in place, after copying the user directory and the journal off the box for the record. Restore flows from version control or a reviewed copy, never from the compromised unit, and re-enter credentials rather than restoring the credentials file. The Node-RED checks tell you where to look first, not whether you're done. Give the boxes with a write path to running equipment the first hour.
Ten minutes in settings.js, then a change request
The vendor's two specific mitigations, apart from updating, are to harden the installation per the Node-RED user guide or to uninstall it, and its standing general recommendation is to protect network access to the device. Hardening is a short list of real option names, and most of it is one file:
adminAuthwith bcrypt password hashes, generated bynode-red admin hash-pw, and per-user permissions. Don't configure the optional anonymousdefaultuser with read access; read access to the editor is a copy of your flows.uiHostset to127.0.0.1, so the listener answers only on loopback. Reach the editor over an SSH tunnel or through a reverse proxy that terminates TLS and authenticates.uiHostmoves the runtime's one HTTP server,http inendpoints and dashboards included, so anything that calls them from another machine needs the proxy route too, and a test before the change window closes.httpAdminRoot: falseon a gateway whose flows are finished. That disables the editor and the Admin API together, and you turn it back on for a change window.httpNodeAuthforhttp inendpoints,httpStaticAuthfor anything served fromhttpStatic. These survive the editor being locked down, and they're frequently the thing a dashboard actually calls.httpswith a key and certificate the site owns, rather than whatever was generated during commissioning. This one stops being optional once the others are set:adminAuthsends a password and then a bearer token,httpNodeAuthandhttpStaticAuthare HTTP basic auth, and over plain HTTP all of them cross the segment in the clear. Either the runtime's ownhttps, or a loopback-bound listener behind a TLS proxy.externalModules.palette.allowInstallset to false, so an editor session can't pull npm packages onto a production device through the palette manager. That switch covers the palette only;externalModules.modules.allowInstallandfunctionExternalModulesgovern what afunctionnode may load, and the shipped settings template leaves the latter on. Set both off, and usenodesExcludes: ['90-exec.js']to drop theexecnode entirely on a gateway whose flows don't need a shell. Search the deployed flows first, forexecnodes and forfunctionnodes with modules on their Setup tab: a flow that references an excluded node type stops every flow from starting. All of this is defence in depth: the function node's sandbox is not a security boundary, so authentication and the network binding remain the controls.- A host firewall on the gateway that admits the runtime's port only from named engineering hosts and from the systems that call its
http inendpoints or dashboards (read the flows for them first), matching ACLs at the conduit, and an audit of the remote-support routers and VPN profiles for forwards or routes that land on it. On a box nobody can touch this week, the conduit ACL is the control that works.
Two mistakes recur in that list. The first is putting a reverse proxy in front of the editor while leaving uiHost at its default, so the hardened route exists and the original one still answers on every interface next to it. The second is treating the proxy's own login as the control; it protects the path through the proxy, and the Admin API on the device is still a POST away for anything on the same segment. Bind the listener first, then decide how people reach it.
Then there's the one that isn't a settings edit. Running the runtime as a service account instead of root means a new userDir, re-checking every file path and serial port permission the existing flows assumed (group membership on the serial device is the one people forget), and a test deploy. It's a small project, it needs a window, and it's the single change that most reduces what a missing login can reach. Treat the vendor version bump the same way: an Industrial OS version is firmware-class change control, not something you push between shifts.
Flows are production software with no home
A flows.json is a deployable artifact. On most sites it gets edited in production, in a browser, by whoever has the tab open, with no review, no diff and no copy anywhere else. Nothing about that is unusual: it's how the tool is meant to feel, and that fluency is why integrators reach for it when a tag has to reach the historian before the line restarts.
IEC 62443 asks you to define zones and the conduits between them. A panel gateway is a textbook conduit, and a conduit whose management interface answers anonymously is one that anyone on the segment can re-plumb. The standard's language assumes you know what's in the conduit. An undocumented flow editor breaks that assumption quietly, long before anyone attacks it, because the asset owner can no longer describe what the device does.
The runtime has an optional git-backed projects mode that keeps each project as its own git repository under the user directory, with commit, diff and remote support; it needs git on the device, and it ships switched off. Turning it on changes the editor's behaviour for the people who use it, which is why it tends not to get turned on mid-project. On a new gateway it costs nothing.
The fix is dull and it works. Keep flows.json and settings.js in version control, deploy them to the device instead of editing on it, review changes the way you'd review a PLC program, and record which gateway runs which revision. That belongs in gateway commissioning, next to the edge hardening that should happen on the same visit, for the ordinary reason that a gateway is as likely to fail as the pump it's watching, and a rebuild at 02:00 shouldn't depend on somebody's memory of a canvas.
What uninstalling costs the people who use it
"Uninstall Node-RED" is sound advice for a device where the flows were a prototype nobody retired. It's the wrong advice for the device where one flow is the only thing feeding the dashboard a shift leader checks every hour, and it's the wrong advice if no one can say which of the two you're holding. That determination takes an afternoon with the flows open, not a line in a remediation plan.
And the ownership question tends to outlast the patch. The flows were often written by an integrator, on a device bought by the automation group, carrying data the process engineers depend on, over a network IT administers. Four parties, no single owner, and a management interface that answered to anyone for however many years the box has been in service. The update to V4.3.4.1 closes one default on one part number. It doesn't assign the box to anybody.
Both candidate fixes sit with someone else. Siemens can keep the service disabled in its images, which does nothing for the copies engineers install themselves. The upstream project could refuse to serve the editor on a non-loopback interface until credentials are set, which would slow the five-minute first run that put Node-RED on plant gateways in the first place. Neither is obviously correct, and this advisory doesn't settle it. So what should a low-code runtime do the first time it starts on an interface that isn't loopback with no credentials configured: refuse, or trust the network?
Notes
Version boundaries in the advisory apply to the IoT2050 Advanced part number only; other SIMATIC families and other vendors' gateways run their own images and need their own check, and a self-installed copy of the runtime has no version boundary at all. Option names above follow the Node-RED 4.x documentation, and older installations may still carry the deprecated httpAdminAuth instead. Locking the editor doesn't reduce what the deployed flows themselves can reach, so not every exposure on these devices is closed by an authentication setting.
References
Reuse & license
This article is published by Zoniax OÜ under a Creative Commons Attribution 4.0 International (CC BY 4.0) license. You are free to share and adapt it for any purpose, including commercially, as long as you give appropriate credit to Zoniax and link back to the original article.
Disclaimer
These Field Notes are general technical information, published as-is for industry peers. They are not professional, engineering, safety, legal, or financial advice, and nothing here is a recommendation to buy, sell, or act. Figures are cited from public sources believed reliable but are not independently guaranteed - verify them against the primary sources and your own plant conditions before acting. Zoniax OÜ and the author accept no liability for decisions made from this content. Naming a standard, product, or vendor is not an endorsement.
Cite this article
Nõmm, A. (2026). CVE-2026-58115: Node-RED's Open Editor on the IoT2050. Zoniax. https://zoniax.com/blog/posts/cve-2026-58115-node-red-open-editor-iot2050
Permalink: https://zoniax.com/blog/posts/cve-2026-58115-node-red-open-editor-iot2050