#91 - January security roundup: CVSS 10 in n8n, self-hosted AI scares, and nonstop patching
One compromised automation host can expose credentials for every service connected to it.
18 min read
We kick off with a CVSS 10 in n8n, then look at self-hosted AI assistants with weak defaults and prompt injection risks. Are your API keys, inbox, and drives safe if a bot is open to the web? What should you rotate, patch, and hide behind a VPN?
The credential pile nobody meant to build
Count the live connections on your automation host. A mailbox, a Drive folder, a database, a CRM, an API key for an LLM provider, plus whatever the sales workflow needed back in November. Each one is a working credential, added on a different Tuesday by whoever needed it that week. Nobody sat down to build a secrets store. One got built anyway, and it has a web login.
January put a price on that pile. CVE-2026-21858, which the Cyera researchers who published it named Ni8mare, carries a CVSS of 10.0 in n8n’s own advisory: no authentication, no user interaction, affected versions >=1.65.0, <1.121.0, fixed in 1.121.0. n8n calls it unauthenticated file access through improper webhook request handling, and the way in is your own workflows: an attacker reaches files on the underlying server “through execution of certain form-based workflows”. Cyera demonstrated it as remote code execution, with an estimated 100,000 servers in range. There is no official workaround, so the fix is the upgrade. Until you can take it, the advisory’s interim step is to restrict or disable publicly reachable webhook and form endpoints, which is the quickest way to put those form workflows out of reach.
Reading that as “another RCE” misses what makes it expensive. Take the instance and you take everything it was wired to, which for many people is a wider set of systems than any production server would hand over. The question isn’t whether you’ve patched. It’s what your host could reach while it ran the older version, and whether anyone wrote that list down. That gap, between how casually these hubs appear and how much they hold, is where the rest of January’s stories start too.
Rotate everything the instance could touch
The cost of a credential hub shows up in the clean-up, so start there.
How much does one automation host actually hold?
More than the list you’d write from memory: API keys, OAuth grants, and for some integrations a plain username and password.
Recovery isn’t “patch and move on”. It’s rotating every credential the instance could reach, and you can’t rotate what you haven’t inventoried. Open the credentials list and write down what each entry can do.
Reissuing API keys is the easy half. An OAuth grant isn’t a password, so revoke it where it was issued rather than assume a password change took it. Google’s rule, set out when it shipped in 2016, revokes tokens on a password change only when the token carries a mail scope, and other providers draw that line elsewhere. Pull the grant from the provider’s connected-apps list, then make the instance ask for a new one.
Scope deserves the same pass. We can’t say how common over-permissioned keys are, and the honest version is a maybe: wire these up quickly, especially without a technical background, and the key that makes the demo work may carry more access than the job needs. So read each grant next to the workflow that uses it.
Rotation is only half the recovery. Exploitation runs through form-based workflows, so ask whether yours had one reachable while it ran a version in >=1.65.0, <1.121.0. If it did, the instance itself is in scope and not only the keys inside it, because Cyera took the same bug as far as code execution: look for workflows nobody wrote, users nobody invited, webhooks nobody registered. Reissuing keys into an instance that still carries somebody else’s workflow is the ending to rule out.
None of which means these tools aren’t worth having. Few teams would hand-build the integrations they ship, which is exactly why the credentials ended up in one place.
Patching Ni8mare is not the same as patching n8n. A second critical advisory landed the same week: CVE-2026-21877, CVSS 9.9, file write to code execution, this one needing an authenticated user. Its fix is 1.121.3, not 1.121.0.
If you can message it, someone else can
The same credential problem arrived in January with a friendlier interface.
Why is a self-hosted assistant riskier than a workflow you wrote?
Because its instructions arrive as text, and a model can mistake instructions in untrusted content for instructions it should follow.
OpenClaw is the example everyone hit at once: an open-source personal assistant you run on your own hardware. Clawdbot became Moltbot on 27 January after Anthropic raised a trademark complaint, then OpenClaw on the 30th. People bought Mac minis to host it, connected it to their mail, calendar and Drive, and added an API key for an LLM provider. The way you talk to it is the product: it lives in your messages.
The project isn’t naive about the first half of that. OpenClaw’s README at v2026.1.29 tells you to treat inbound DMs as untrusted input, and ships a pairing policy to back it: an unknown sender gets a short pairing code and, in the README’s words, “the bot does not process their message” until you approve it. Channels carry their own allowlists, channels.whatsapp.allowFrom and the like, and opening the assistant to the public takes a deliberate dmPolicy="open" with "*" in it. If the worry were only that a stranger might DM your bot, config answers that.
The worry doesn’t stop there, and this is the part an allowlist can’t reach. It decides who may start a conversation, not what the assistant reads once it’s working. Ask it to summarise your inbox and it ingests an email somebody else wrote; ask it to check a page and it ingests whatever is on the page. The hostile instruction arrives as content, inside a session you opened yourself, from somebody who was never going to be on the list.
That’s the gap underneath the authentication. Pairing settles who opened the conversation; the text it reads afterwards carries no such marking, and an instruction looks like an instruction whoever wrote it. So we’d put it bluntly: as far as we could tell, there’s no way today to expose an LLM connected to valuable data to the public and keep it safe. Nor is the problem limited to text. Steering an autonomous system with a printed sign its camera reads as an instruction is an older trick, and the lesson transferred intact.
Publishing the port is the decision
Self-hosting is only private until something answers on a public address.
How long before someone finds the host you just put online?
There is no safe answer to plan around, because mapping the address space is somebody’s business model.
Shodan and Censys scrape the internet, fingerprint what answers, and publish or sell the result. Censys’s own account of its scanning puts 137 IANA-assigned ports across the whole IPv4 space on a daily cycle, a further 3,455 ports on a ten-day walk, and the rest at a low background rate. Read that as coverage, not as a countdown: it says how often one scanner sweeps which ports, not how long any host stays unseen. A common port gets catalogued quickly, and an odd one can be found just as fast.
The numbers moved fast here. Censys counted 21,639 exposed OpenClaw instances on 31 January, up from roughly 1,000 in under a week, matched on the control panel’s HTML title. Most still asked for a token, which is the detail worth sitting with: reachable is step one, and step two is whichever CVE lands next. The routine after a disclosure is mechanical: pull the addresses running the affected version, start trying.
Credit to the project for moving. OpenClaw’s v2026.1.29 release on 30 January removed gateway auth mode none outright and made the default fail closed. Fixed defaults don’t reach back to instances installed under the old ones, and the install instruction people pasted into a terminal was one line.
The tool itself is good. Run it in your own Kubernetes cluster with its access scoped down and it earns its keep. Scoped down means its own accounts rather than yours: a calendar server stood up for the bot alone, a mailbox it may read and not much else. An assistant holding your mail belongs behind a VPN or a tailnet. The README’s remote route binds the gateway to loopback and reaches it over Tailscale Serve or an SSH tunnel, which keep it inside your tailnet or your own machine. Funnel is on that list too and is a different proposition: the README calls it public HTTPS, so loopback binding and a password become your whole perimeter. That is presumably why Funnel refuses to start unless gateway.auth.mode is password.
The patch stream nobody writes about
A CVSS 10 is a headline. The baseline underneath it isn’t.
If you only have attention for one thing, should it be the headline CVE?
No, because the headline CVE is the one your team has already heard about.
Microsoft’s own January 2026 release notes run to more than 110 CVEs, eight of them rated Critical. That is one month from one vendor, and it repeats every month, while SAP, Cisco and the rest run their own cycles on their own calendars. Almost none of it is news.
Minor releases are where this goes wrong. Teams schedule majors and skip patch releases as not worth the disruption, which is backwards: the patch release usually exists because something was found.
What is better than a few years ago is visibility. Dependabot and repository scanning tell you what you’re running and what’s behind, where you used to track it by hand or not at all.
The app you can’t refuse
Patching assumes the software is yours to patch. A growing amount of it isn’t.
What do you do when somebody else’s security is the price of your kid playing football?
Very little, which is exactly why the obligation has to sit with the vendor and with the regulator.
Sweden’s data protection authority fined Sportadmin SEK 6 million, roughly EUR 560,000, on 28 January for breaching Article 32 of the GDPR. The attacker reached data on more than 2.1 million individuals, which IMY says “mainly concerned children and young people”: names and contact details, personal identity numbers, sport and club, along with sensitive health data and, to some extent, data on people with protected identity. The attack was in January 2025 and the data ended up on the darknet. IMY found no system to detect intrusions in real time, and recorded that “for a long time prior to the attack, Sportadmin was aware of certain weaknesses in its systems and of areas with elevated risks of attack”.
Parents in those clubs had no way out. The schedule is in the app, the matches are in the app, the messages are in the app. Refusing it means your kid misses training.
We won’t pretend a fine creates engineering capacity at a small vendor, and prioritising features over hardening is what most teams do until something forces the question. Known risk plus no detection is a different category, and that’s where the regulator landed. There’s an arithmetic problem too: a personal identity number can’t be rotated by its owner, and the data is worth something to whoever buys it for phishing. If the penalty lands below what the data fetches, the incentives don’t move.
Who else can prove that they’re you
Software you can’t refuse shows up at the government counter too, where the credential is the whole system.
What’s wrong with a login certificate you can export?
If you can copy it, you can give it away, and people do.
In Spain the certificate you use to identify yourself to the tax agency and sign documents installs into your browser and exports straight back out, private key included. It’s normal practice for accountants to ask clients to send the certificate over, and clients send it. The correct path is to register representation so the accountant acts under their own identity, and Spain’s tax agency documents an online route for that, granted with a DNI electrónico, an electronic certificate or Clave PIN. The alternative is the counter, a visit and an afternoon, which is what the shortcut competes with.
Compare the Nordic shape. Sweden runs BankID: the bank sends an authentication request, you approve it in the app on your phone with a biometric or a passcode, BankID confirms to the bank that it was you. Denmark runs a separate system, MitID, described by the Danish Agency for Digital Government as “primarily an app where you can approve actions online with a swipe”. Two systems, one shape: a third party in the loop, approval per action on a device you’re holding, and nothing you can export and mail to an accountant.
The second half of the Spanish story is about trust stores. AutoFirma, the desktop signing client that uses your certificate, doesn’t hold it: its installation manual says the keys stay in your operating system or browser key store, or on a smart card, and never leave. What AutoFirma does install is a certificate of its own, added at setup to the system and Firefox trust stores so the browser can reach the app over a local socket. We noticed sideways: a municipal site in Las Palmas loads clean on a machine with AutoFirma installed and warns of an untrusted certificate on one without. We haven’t confirmed which CA issues that site’s certificate, so treat the chain as unverified. The general rule doesn’t depend on it. An unconstrained root that your store trusts for TLS can authorize a certificate for any domain, your bank’s included. Constraints can narrow that: RFC 5280 lets a relying party apply name constraints to a trust anchor. That’s also why root programs are strict: when Kazakhstan pushed a government root onto its citizens in 2019, Firefox and Chrome shipped measures to block it.
Credentials leak from laptops, not from the exchange
Wherever the identity model ends up, it ends up on an endpoint.
Where do the credentials in these dumps come from?
From infostealer malware on people’s own devices, not from the platforms whose logins show up in the dump.
On 23 January, researcher Jeremiah Fowler reported an unprotected database of 149,404,754 logins and passwords, about 96 GB, which his report describes as keylogging and infostealer malware data, not one platform’s breach. Cointelegraph put about 420,000 of those records against Binance accounts and got Binance’s answer: infostealers take credentials from compromised devices, so “those are not leaks from Binance”. Self-serving, but it matches the collection method Fowler documented.
Clean the machine before you reset anything. Fowler’s recovery guidance is explicit about the order: while the malware is still running, your replacement password gets captured like the old one. Isolate the device, scan it with current security software, rebuild if anything turns up, then reset credentials and revoke sessions from a machine you trust.
Be precise about what a stolen password buys. It gets the attacker to the login page with the right answer, so whatever guards that page still does: a second factor, a device check. Stealing a session cookie is the separate technique: malware lifts an already-authenticated session and replays it without touching the login. A credential dump is evidence of the first.
That changes what a second factor should be: ask for phishing resistance, which NIST’s SP 800-63B requires at its top assurance level. A FIDO credential is scoped to a relying party id, a domain name: WebAuthn lets a page ask only for credentials under its own domain, and the service checks the origin in the signed response before accepting it. A lookalike domain fails both tests. One-time codes carry no such scoping, and NIST doesn’t permit them at that top level. So make the backup a second FIDO authenticator, not a code generator: an account that still accepts a code keeps a phishable door. Passkeys are the usable form of FIDO, but NIST notes that syncable authenticators “require the private key to be exportable” and rules them out there. Where money is involved, pick device-bound.
Biometrics sit in the middle. NIST is blunt that “a biometric characteristic is not recognized as an authenticator by itself”. Under WebAuthn the verification happens locally: what reaches the service is a signed assertion with a user-verification flag, not your fingerprint. The key proves, the finger only unlocks it.
What this means for teams
The thread this month isn’t a clever attack. It’s that credentials have quietly accumulated in places nobody treats as a credential store: an automation host, a personal assistant, a browser profile, a signing client. Four things worth doing this week:
- Write down what your automation host and any assistant can reach, every API key, OAuth grant and mailbox, then reissue what’s over-scoped at the narrowest scope the workflow needs. Revoke OAuth grants at the provider rather than trusting a password change to do it.
- Upgrade self-hosted n8n past both January advisories, not just the famous one: Ni8mare is fixed in
1.121.0, the authenticated file-write RCE in1.121.3. While you wait for the window, take public form and webhook endpoints out of reach, and give every self-hosted tool a named owner who reads its advisories and applies its patch releases, not only its majors. - Check what of yours answers on a public IP. Anything agentic goes behind a VPN or a tailnet, bound to loopback, with gateway auth required, its sender allowlist set rather than left open, and its own accounts rather than yours.
- If one of your machines turns up in a credential dump, clean or rebuild it first, because a new password set from an infected device gets stolen too. Then put a phishing-resistant factor on your mail and on anything holding money, register a second one as the backup so you can switch the one-time codes off, and spend twenty minutes doing the same for one family member.
Common questions, answered
How do I run a self-hosted AI assistant without exposing my email?
Keep it off the public internet: bind the gateway to loopback, require token or password auth, and reach it over a VPN, a tailnet or an SSH tunnel rather than a public tunnel like Tailscale Funnel. Set the sender allowlist rather than leaving it open, which stops strangers starting a conversation. That still leaves injection: hostile instructions ride in on content the assistant reads for you, an email body or a fetched page, not only on messages sent to it. So give it its own accounts at narrow scopes instead of your personal credentials.
What should I rotate after an automation platform is compromised?
Every credential that instance could reach, starting with the ones that can spend money, send mail or read customer data. Rotating API keys isn’t enough: revoke OAuth grants at the provider too, since a password change may leave them working, and go through the instance for workflows, users and webhooks nobody on your team created. Reissue each credential at the smallest scope its workflow needs, which is the only part of this that pays off twice.
Is fingerprint unlock enough for an account with money in it?
No. Register a FIDO credential on the account itself, a security key or a passkey, and prefer a device-bound one over a synced one where money is involved. Your fingerprint is checked on your device and the service never receives it; what it verifies is a signature, with a flag saying verification happened, so the biometric counts only when a key bound to that account sits behind it. Treat the device as part of the account: a password stolen by malware still has to get past whatever guards the login, while a stolen session cookie skips the login entirely.
Related episodes
- Episode 80, Understanding Passkeys: Benefits and Limitations is the model this post leans on: what binds a person to a device, and what passkeys don’t fix.
- Episode 81, Keeping Secrets Safe covers where secrets belong per consumer, and what makes rotation possible instead of theoretical.
- Episode 82, Tools, MCPs and Attack Scenarios is the long version of the injection argument: how tool use gets turned against the person running it.
- Episode 86, MCP plugins: your next security blind spot? works through the blast radius of connected tooling, the same question an automation hub poses.
Resources
- n8n advisory GHSA-v4pr-fm98-w9pg (CVE-2026-21858) settles the CVSS 10.0 rating, the affected range
>=1.65.0, <1.121.0, the fix in1.121.0, exploitation through form-based workflows without authentication, and the interim step of restricting public webhook and form endpoints. - n8n advisory GHSA-v364-rw7m-3263 (CVE-2026-21877) settles the second January advisory: CVSS 9.9, authenticated file write to code execution, fixed only in
1.121.3. - Ni8mare, unauthenticated RCE in n8n, Cyera settles the remote code execution demonstration and the estimate of 100,000 affected servers.
- OpenClaw README at
v2026.1.29settles the January defaults: DMs as untrusted input, pairing for unknown senders, per-channel allowlists, loopback binding, and which remote routes are private. - OpenClaw in the Wild, Censys settles the count of 21,639 exposed instances on 31 January 2026 and how they were identified.
- Administrative fine against Sportadmin, IMY settles the SEK 6 million fine under Article 32, the 2.1 million individuals affected, the data categories, and the finding on known risks.
- 149M logins and passwords exposed online, Jeremiah Fowler settles the record count of 149,404,754, the 96 GB size, that the database held infostealer malware data rather than one platform’s breach, and why cleaning the device comes before changing passwords.
- NIST SP 800-63B, Digital Identity Guidelines settles where phishing resistance is required, why one-time codes don’t qualify, and why syncable authenticators are ruled out at the top level.
Since we recorded
OpenClaw’s creator, Peter Steinberger, announced in February 2026 that he was joining OpenAI, with the project moving to a foundation to “stay open and independent”. Version 2.0 shipped in August 2026 with an opt-in secrets proxy: protected secrets stay out of the text the model sees, and the gateway inserts them into requests to approved destinations. The substitution covers supported gateway-hosted HTTPS commands that go through the proxy, so other execution paths fall outside it, and it narrows what an injected instruction can walk away with rather than whether the injection lands. We came back to the wider question in Episode 111, AI Agents: When Helpful Becomes Harmful, on what an agent holding your credentials does with an instruction it reads too generously.
