9.4
CVSS 4.0 base score, per Perforce
1
CA-signed agent certificate is all the attacker needs
14 days
from the fix on 3 Nov to the KEV listing
1 Dec
federal remediation deadline set by CISA
The certificate authority is the perimeter
CVE-2026-115731 is described everywhere as an authenticated flaw, and that word is doing more work than it should. The authentication is an agent certificate issued by the Puppet server's own certificate authority, and in a large share of real deployments that CA signs whatever asks it nicely.
Puppet is the control plane for a fleet. The server compiles the catalog that tells every managed machine which packages to install, which users exist, which files to write and which services to run, and the agents apply it as root on the next run. Code execution on that server is not one compromised host. It is a position from which every managed host can be told to do something, by the system whose whole job is to be obeyed.
So the useful question this week is not only "have we patched the Puppet server?" It is "who can get an agent certificate from our CA, and how would we know if someone did?" On a tightly run deployment the answer is a short list of provisioning systems and a human who signs requests by hand. On many we see, it is anyone who can reach the server and type a hostname that matches a glob. That second group can now run code on the server.
The flaw is fixed. The trust model that turns it from an insider problem into an outsider problem is a configuration choice, and it will still be there after the upgrade unless someone goes and looks.
What CISA and Perforce say about CVE-2026-115731
On 17 Nov 2026 CISA added CVE-2026-115731 to its Known Exploited Vulnerabilities catalog, based on evidence of active exploitation, under the name Perforce Puppet Enterprise Deserialization of Untrusted Data Vulnerability. The catalog entry gives federal civilian agencies until 1 Dec 2026 to apply the vendor's mitigations, and lists the weakness as CWE-502 and CWE-94.
The flaw lives in the report processing code of Puppet, as used by Puppet Server. In the words of the Perforce advisory, a client that holds any CA-signed agent certificate can execute arbitrary code on the Puppet Server as its service user by submitting a report containing a crafted rich-data type through a PUT request. The same sentence carries the qualifier that matters: where the certificate authority autosigns requests without strong verification, an unauthenticated network attacker can obtain such a certificate.
Perforce scores it 9.4 under CVSS 4.0, critical, network reachable, low complexity, low privileges, no user interaction, with high impact on the server and on subsequent systems. NVD scores it 9.9 under CVSS 3.1 with a changed scope. The fixes shipped on 3 Nov 2026, so the KEV listing came 14 days after the patch.
-
Product //
Puppet Server in Puppet Enterprise and Puppet Core; OpenVox Server; Red Hat Satellite -
Affected //
PE 2026.0.0; PE 2025 before 2025.12.1; PE 2023.8 before 2023.8.13; Puppet Core 9 before 9.1.3 and 8.x before 8.22.2 -
Fixed in //
PE 2026.0.1, 2025.12.1, 2023.8.13; Puppet Core 9.1.3, 8.22.2; OpenVox Server 9.2.0, 8.18.0 -
CVSS //
9.4 (4.0, Perforce); 9.9 (3.1, NVD) -
Needs //
one CA-signed agent certificate and network access to the server -
Exploited //
Yes, per CISA KEV, added 17 Nov 2026
The upstream picture is wider than the KEV name suggests. The OpenVox project, the community fork of open source Puppet, announced on oss-security the same day that it had fixed two vulnerabilities rated above 9 that affect all known OpenVox and Puppet versions and implementations, with openvox-server 9.2.0 and 8.18.0 carrying the server fix. Its explanation of the bug is the clearest public one. The report is a JSON document the agent sends after every run, listing what changed. For some custom resource types an event inside that report can carry a serialized Ruby object rather than plain JSON, and when the server processes the report it turns those objects back into live ones. A malicious event object can be built so that code runs on the server when that happens.
The second OpenVox fix is a client impersonation flaw in trapperkeeper-authorization, the library that enforces auth.conf. Plain string rules were matched with wildcard logic, so a certificate named evil.target.example.com could act as target.example.com. It is a separate bug, but it lives on the same trust boundary: the server deciding what a certificate is allowed to be.
Red Hat rates the issue critical because the service account it runs as can compromise the central certificate authority and all downstream managed systems, and it notes that Satellite processes incoming client reports over the network by default. Red Hat shipped fixed openvox-server packages for Satellite 6.16, 6.17 and 6.19 on 5 Nov 2026.
On attribution the record is thin, and we will not fill it in. CISA has not said who is exploiting the flaw or how it came to know. Security Affairs reports that Perforce has said it is not aware of attacks exploiting the flaw and that no threat actor or campaign has been linked to it. Both statements can be true at once: a vendor sees its customers' support cases, and CISA sees incident reports from elsewhere. The KEV listing is the operative fact for prioritisation.
How one agent report becomes control of the fleet
The chain is short because Puppet is built to make it short. Agents are supposed to talk to the server, the server is supposed to accept their reports, and the server is supposed to be trusted by every agent. The flaw only has to connect those facts.
Figure 01
From a signed request to every managed node
- 01
Enrol //
Get an agent certificate, legitimately or through loose autosigning.
- 02
Report //
Submit one crafted run report to the server.
- 03
Execute //
Code runs on the server as its service user.
- 04
Hold //
Reach the CA keys, code and secrets in the same process.
- 05
Push //
Every agent applies what the server tells it to.
Result //
One certificate became the configuration of the estate.
Step one is where the deployment, not the bug, decides the outcome. An attacker already inside a managed machine has a certificate on disk and needs nothing else. An attacker with only network reach needs the CA to sign a request for them, which is the autosigning question in the next section.
Step four is why Red Hat talks about the CA. Puppet's own documentation describes Puppet Server as a Ruby and Clojure application in a single JVM process that runs the catalog compiler in JRuby interpreters and also provides the certificate authority. Code running as that service user is running beside the CA's credentials, the Puppet code that defines every node, and whatever secrets that code pulls in for catalogs: database passwords, API tokens, keys for the cloud accounts the fleet runs in. An attacker who holds the CA key can mint agent certificates at will, which means rebuilding the server is not enough to evict them unless the CA is rebuilt too.
Step five is the property defenders forget. The agents run as root and do what the catalog says. Nobody needs to move laterally in the usual sense when the system that writes configuration is already reachable from every host on the next run. We made the same point about remote management tools in our N-central write-up: management planes are the highest value targets in an estate, because compromise of the plane is compromise of whatever it manages.
Why autosigning decides who is authenticated
Out of the box, the Puppet CA does not sign anything on its own. The documentation says that certificate requests must be signed manually by an admin, with the puppetserver ca command or the Node requests page in the Puppet Enterprise console. Autosigning exists to speed up bringing new nodes into the deployment, and the same page warns that it changes the nature of the deployment's security.
There are three ways it gets switched on, and they are not equal.
Naive autosigning sets the autosign setting to true and signs every request. On a server reachable from user networks, that makes the advisory's "authenticated" mean nothing.
Basic autosigning uses autosign.conf, an allowlist of certnames and domain globs such as a whole internal domain. Puppet calls it insecure in plain words, because any host can provide any certname when requesting a certificate, and it states that an attacker who guesses an unused certname allowed by the file can obtain a signed agent certificate and pull a catalog. The certname is just a string the requester chooses. A glob over the internal domain is an open door for anything on the network that can reach the CA.
Policy based autosigning runs an executable for each request and signs only when it approves. Puppet's guidance is careful here: it can be extremely secure, for example when each node carries a unique pre-shared key, but it is only as good as the secret or the system it checks. A policy that checks a cloud instance identifier is only as strong as the cloud API; anyone who can create a rogue instance in the account can get a certificate.
This is why we read the advisory's own qualifier as the headline. The bug is a deserialization flaw with a fix. The exposure is the population of machines and people who can hold a certificate, and that population is set by an autosign choice somebody made years ago to make provisioning faster. Fleets that were built when every new virtual machine needed to enrol without a human usually still have the shortcut, long after the provisioning system that needed it was replaced.
The OpenVox project's own mitigation list for anyone who cannot upgrade starts in the same place: reach the server only from the network segments that need it, and disable or tighten autosigning. Red Hat's mitigation text says the same about autosigning without strong verification.
How to detect exploitation on a Puppet server
Start from the trust boundary rather than the payload. Two questions cover most of it.
First, which certificates has the CA signed, and do they all belong to machines you provisioned? The puppetserver ca command lists signed certificates. Compare that list, and especially anything signed since mid October when the flaw was reported upstream, against your inventory or CMDB. A certname that matches the naming pattern but no machine you own is the finding. Revoke it, and treat the server as suspect until the next question is answered.
Second, did the server process do anything a catalog compiler should not? Puppet Server is a long running JVM under the puppetserver service. Its normal children are few and predictable for a given deployment: an autosign policy executable if you use one, an external node classifier, perhaps a Hiera backend script. A shell, a download tool or a scripting interpreter launched from that process outside those is worth a look on any day, and this month it is the trace that code execution through a report would most plausibly leave.
Figure 02
Puppet Server spawns an unexpected child process
Sigma
title: Unexpected child process of Puppet Server
status: experimental
description: A shell, interpreter or download tool started by the Puppet Server JVM, the likely trace of code execution through report processing (CVE-2026-115731)
logsource:
product: linux
category: process_creation
detection:
parent:
ParentImage|endswith: '/java'
ParentCommandLine|contains: 'puppetserver'
child:
Image|endswith:
- '/sh'
- '/bash'
- '/dash'
- '/curl'
- '/wget'
- '/python3'
- '/perl'
- '/nc'
filter_known:
CommandLine|contains:
- '/etc/puppetlabs/puppet/autosign'
- 'external_node'
condition: parent and child and not filter_known
falsepositives:
- Autosign policy executables, external node classifiers and custom functions that shell out by design
level: high Two more places are worth searching. The OpenVox advisory describes the trigger as a rich-data type inside a report event, carried in the event's type field. If your deployment stores reports or forwards them to PuppetDB or a reporting tool, look for events whose type is an object definition rather than a plain value, especially from certnames that do not usually run custom types. Legitimate custom types can produce rich data, so treat a hit as a lead to triage, not a verdict.
Then look at the server's outbound traffic. A Puppet server talks to its agents, its database and its code repository. A new outbound connection to an address on the internet from that host deserves the same urgency as a domain controller doing the same. For the broader shape of that review on an internal network, our Learn guide on internal Active Directory recon walks through how we map what a host should and should not reach.
How we test configuration management on an engagement
On internal engagements the configuration management server is one of the first things we look for, because it answers the question every assessment is really asking: if an attacker gets a foothold, how far can one system carry them? Puppet, Satellite, Ansible controllers, SCCM and RMM platforms all sit in that class.
At a high level, on a sanctioned engagement, we work through the same trust boundary this post describes. Can a machine we control on a user or guest network reach the server at all? If it can, will the CA sign a request for a name we choose, and does the allowlist or policy check anything an attacker could not supply? If we hold a legitimate agent certificate from an existing host, what does the server give us without any bug: the catalog for that node, the secrets inside it, and through auth.conf, anything belonging to other nodes? None of this needs an exploit to produce a serious finding, and in our experience the autosign and auth.conf review finds more than the version check does.
Where the scope allows it, we then show impact without damage: proof that a certificate was issued to a host the client does not own, or that a catalog with secrets was readable, rather than pushing a change to the fleet. The report then rates the finding by what the server controls, not by the CVSS of whichever bug happened to be current that week. A low severity "autosigning enabled" line in a scanner export and a critical "any host on the guest network can enrol as a production web server" are the same configuration described honestly.
What is still coming
Three things are worth watching.
The full technical write-up. The OpenVox project credits the researchers at Neodyme with the report and says their write-up will follow. A detailed public account usually widens exploitation, so assume more scanning of Puppet servers once it lands, not less.
Repackaged copies. OpenVox says the bug affects every known Puppet implementation, and it asks anyone who bundles or repackages OpenVox to update too. That means Puppet servers inside appliances, distribution packages and lifecycle tools like Satellite, which do not show up in an inventory search for "Puppet Enterprise". The OpenVox project also states that open source Puppet will not receive further updates, so a server still running an old open source Puppet package from a distribution needs a migration plan, not a patch.
The impersonation fix. Teams with custom auth.conf rules should check them after the upgrade. The OpenVox project warns that some custom rules may have depended on the old wildcard behaviour and may need updating, and a rule that silently stops matching is its own outage.
The wider lesson will outlast this CVE. Every fleet tool has an enrolment path, and the enrolment path is the real perimeter of the control plane. Puppet calls it autosigning. Other tools call it registration tokens, bootstrap keys or enrolment profiles. Whatever the name, if it was configured for convenience, it is worth an hour this month.
What to do today
-
Upgrade every Puppet server. 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; the 5 Nov Satellite errata. Include appliances and compile servers, not only the primary.
-
Turn off naive autosigning and replace glob allowlists. Sign by hand, or use a policy executable that checks a secret an attacker cannot supply.
-
Audit what the CA has signed. List every signed certificate, reconcile it against inventory, and revoke what you cannot account for, starting with anything signed since mid October.
-
Restrict who can reach the server. Only the networks that host managed nodes, never user, guest or partner segments.
-
Hunt on the server itself. Run the Sigma rule above, review outbound connections, and if anything looks wrong, plan to rebuild the CA as well as the host.
-
Review auth.conf after the upgrade. Confirm custom rules still match what you intended.
SOURCES
- CISA // adds CVE-2026-115731 to the KEV catalog on 17 Nov 2026 based on evidence of active exploitation
- Perforce advisory // affected and fixed Puppet Enterprise and Puppet Core releases, CVSS 4.0 score of 9.4
- OpenVox on oss-security // how the report event deserialization works and the fixed OpenVox releases
- Red Hat // Satellite ships the affected server, and the CA and managed systems are at risk
- Puppet documentation // basic autosigning lets an attacker who guesses an allowed certname obtain a signed certificate