Back to Blog

Puppet's patch does not revoke what your CA already signed

Communications cabinet with patch panels and hundreds of unlabelled blue and grey network patch leads

The CVE-2026-115731 entry in CISA's KEV catalogue has two sentences. Everybody quotes the first: a client holding any CA-signed agent certificate executes code on the Puppet Server. The second has barely been read, and it is the one that decides whether this is bad or very bad: "If the certificate authority automatically signs requests without strong verification, an unauthenticated network attacker can obtain a valid certificate and exploit the vulnerability." Translated: your severity is not set by the CVE. It is set by one line of your puppet.conf.

Perforce published the advisory on 3 November. CISA added it to the Known Exploited Vulnerabilities catalogue on 17 November, with a US federal deadline of 1 December. Between the two there were fourteen days in which this was, for almost everybody, another 9.4 you had not heard about. Us included.

What exactly breaks

Puppet works like this: every managed machine runs an agent, the agent asks the server for its catalogue, applies it, and then sends a report back describing what changed. That report travels as a PUT to /puppet/v3/report. The report-processing code accepts Puppet's rich data types, and that is where the hole is: a report crafted with a malicious rich-data type gets deserialised and executes arbitrary Ruby code on the server, with the service user's permissions. CISA files it under two weaknesses at once, CWE-502 (deserialisation of untrusted data) and CWE-94 (code injection). As for the score, everyone gives their own: Puppet rates it 9.4 under CVSS 4.0 and the NVD 9.9 under CVSS 3.1. The KEV publishes none.

Note the direction of the traffic, because that is what makes this different from any old flaw in an admin panel. Nobody comes in through the front door here. They come in through the channel you built so that your own machines could report back how they got on. The agent holds a certificate because it needs to say "I am web-07". What nobody ever decided is that the same certificate should also be good for running code on the machine that configures all the rest.

The line that sets your severity

Go back to the KEV's second sentence. Without autosigning, the attacker needs an already-issued certificate: they must first have compromised some machine of yours — any of them, the dumbest one. With autosigning and no strong verification, you hand it over: they request a certificate, your CA signs it by itself, and with that they go and execute code on the server. Same CVE, two completely different stories.

And here is the uncomfortable part, which is not us talking: Puppet has been saying it in its own documentation since long before this CVE existed. On naïve autosigning, the familiar autosign = true: "For security reasons, never use naïve autosigning in a production deployment." And on the allowlist variant in autosign.conf, which plenty of people take to be the grown-up version: "Basic autosigning is insecure because any host can provide any certname when requesting a certificate."

So the CVE did not create this exposure. It collected on it. For years the cost of leaving autosigning open was theoretical, while the convenience of provisioning machines without approving anything by hand was real and daily. On 3 November the cost stopped being theoretical. Checking which rung you are on is one command: puppet config print autosign --section server.

What the patch does not do

This is the part that made us write the post, and it is where we part company with how the story has been told. It has run everywhere as "patch now", and fine, patch now. But the Perforce advisory says something more that almost nobody picked up, and the whole sentence is worth reading: a successful attack "could expose the server's certificate-signing key, allow an attacker to issue trusted certificates, and enable changes to software and configuration delivered to managed systems on their next Puppet run". That third clause is the one that usually gets cut, and it is the one that actually describes the reach.

Read it slowly. On a Puppet Server acting as its own CA — the default arrangement, and the most common one — the certificate authority's private key lives on the very machine where, if anyone did exploit this, arbitrary code ran. The service user can read it. And an apt upgrade only goes so far there: that key will sign fresh certificates whenever the attacker feels like it, under names you would recognise, and the patch does not revoke a single one of them. It shuts the door that was used; the lock that would have walked off is still the same lock.

This is not the first time we have written about the same thing with a different name on it. A few days ago it was the CephX fix, where the problem was likewise a system trusting a credential it had itself validated badly, and the operational half of the fix — rotating keys, moving the daemons up a version — was not in the headline either.

The three rungs, and why nobody has rehearsed the third

This gets answered in three rungs, and they go in that order. Anyone can do the first two in an afternoon. The third is the one almost nobody has written down anywhere.

Rung What you do What it rules out
1 Move to a fixed version: Puppet Enterprise 2026.0.1, 2025.12.1 or 2023.8.13; Puppet Core 9.1.3 or 8.22.2; OpenVox Server 9.2.0 or 8.18.0 (with agent 9.3.0 or 8.32.0). Future exploitation. Nothing in the past.
2 Check your autosign setting (puppet config print autosign --section server), list every certificate (puppetserver ca list --all) and look for names you do not recognise. And in the server access log, PUT /puppet/v3/report from certificates that are not agents of yours. Doubt. You leave knowing whether you need rung 3 or not.
3 Only if there is evidence: rebuild the server, regenerate the CA and reissue every certificate, and rotate the secrets your catalogues hand out. The attacker who already holds your signing key. It is the only rung that does.

Rung 3 is in nobody's plan, and there is a reason. Puppet's own documentation introduces it with this warning: "This process destroys the certificate authority and all other certificates. It is meant for use in the event of a total compromise of your site, or some other unusual circumstance." A stolen signing key is that circumstance. But while you do it, every agent in the estate stops talking to the server until it re-enrols, and that is not ten minutes: it is the whole estate, including the machines that happen to be switched off that afternoon and the ones belonging to the customer who does not answer the phone on a Saturday. And some of them will be holding a certificate issued by somebody who left years ago.

Who signs off on stopping

With the consoles that command the whole estate, whatever the brand, we see the same thing again and again: they end up with no clear owner. Security reads the KEV and raises the ticket; infrastructure owns the box but did not write the manifests. And whoever does automation did write them, but does not touch production. The rung-3 decision — stopping the estate for a while to rotate a CA, on evidence that is not yet certainty — sits in the gap between the three.

That is why our RID Programme puts security, workplace, infrastructure and data into a single team with shared goals, rather than four suppliers passing the parcel. We do not say that as an abstract virtue: in a case like this, somebody has to be able to say "we stop and we rotate" without calling three meetings. And following the KEV catalogue and deciding what in it actually applies to you falls under managed cybersecurity — unglamorous work, and the sort that heads off the most trouble.

Two ways to misread this

The first: if your Puppet is a throwaway lab, autosign = true is fine and rotating the CA is theatre. But mind the exact condition, which is narrower than it looks: Puppet's documentation reserves naïve signing for temporary test deployments incapable of serving catalogues containing sensitive information. It is not enough that it happens not to serve secrets today; it has to be unable to. If it clears that bar, upgrade and get on with your life.

The second, and this one matters more: none of the above is an argument against configuration management. Take away the server that configures the estate and what you are left with is hand-built machines, each one different, none documented, and you find out on the day they all need patching. We have seen that, and it is considerably worse. Concentrating control carries a real risk, and the answer is knowing where it is concentrated and treating that machine accordingly. It is the same reasoning we applied when an RMM vendor had to ask its customers to switch off the console that manages every machine.

And one more piece of honesty: we do not know how many people have been exploited through this. CISA records the date something enters the catalogue, not a victim count, and presence in the KEV means known exploitation, not mass exploitation. What should be said, because it cuts the other way, is that the public OpenVox advisory of that same 3 November includes a complete proof-of-concept report that runs a shell command. Which means the route has been published since day one. The cost of checking whether this touches you is an afternoon; the cost of not checking is on you.

So we close on a question that either gets answered in ten seconds or does not: how many certificates has your Puppet CA signed, and how many of those machines still exist? If you cannot work out the difference between those two numbers, 115731 did not bring you the problem. It showed it to you.

Sources (verified at origin): the vulnerability description, the two sentences quoted from the catalogue, the CWE-502 and CWE-94 weaknesses, the date added (17/11/2026) and the due date (01/12/2026) come from CISA's KEV catalogue, downloaded as its JSON file on 19/11/2026. The 9.4 is the CVSS 4.0 score published by Puppet itself; the 9.9 is the NVD CVSS 3.1 score, which also confirms the 03/11/2026 publication date. The KEV publishes no score. The fixed versions of Puppet Enterprise, Puppet Core and OpenVox, and the full sentence about the certificate-signing key — Perforce advisory, also reported by SecurityWeek. The two autosigning quotes — "never use naïve autosigning in a production deployment" and "basic autosigning is insecure…" — and the warning that regenerating the CA "is meant for use in the event of a total compromise of your site" are from Puppet's official autosigning documentation and its CA regeneration procedure. The fourteen-day arithmetic, the three-rung ladder and the organisational reading — whose server the configuration master is — are ours. That the public OpenVox advisory of 03/11/2026 includes a complete proof-of-concept report running a shell command — Ostorlab analysis. Photograph: "Cable closet", public domain via Wikimedia Commons, cropped. We are not a Perforce or Puppet partner or reseller, and we do not sell anybody's licences.

Who owns your configuration server?

At everyWAN the KEV alert, the upgrade and the decision to rotate all sit with the same team, not three different suppliers. That is the idea behind the RID Programme, together with our managed cybersecurity. If you do not know how many certificates your CA has signed, write to us and we will count them together.

Talk to everyWAN

Tags:

Share:

Subscribe to our newsletter

To receive IT stories, everyWAN news and exclusive subscriber offers, sign up to our mailing list

Minorisa de Sistemas Informaticos y Gestión S.L. © 2026
everyWAN
everyWAN