Every Puppet agent sends a report to its server after a run finishes, and the report holds logs, metrics and events. The agent PUTs that JSON to /puppet/v3/report. Some events carry a serialized Ruby object instead of plain data, and the server deserializes them back into something the compiler can use.
The OpenVox advisory traces it from there. The JSON lands in Puppet::Pops::Serialization::FromDataConverter.convert, the __ptype field names a type, Types::TypeParser parses that name, and RubyGenerator.create_class builds a class and hands it to class_eval, where a backtick in the annotation body runs at eval time 1. The advisory publishes the full JSON event that does it 1.
Perforce scored CVE-2026-115731 at 9.4 critical under CVSS 4.0 2. The NVD record is marked Analyzed and carries a CVSS 3.1 primary score of 9.9 critical from nvd@nist.gov, with a vector of AV:N/AC:L/PR:L/UI:N/S:C 3. The two bodies scored the same bug on different scales, so I keep both numbers and name the body behind each.
Puppet Enterprise 2023.8.0 before 2023.8.13, 2025.0.0 before 2025.12.1, and 2026.0.0 before 2026.0.1 carry the flaw, and 2023.8.13, 2025.12.1 and 2026.0.1 close it. Puppet Core 9.0.0 before 9.1.3 and 8.x before 8.22.2 carry it as well, and 9.1.3 and 8.22.2 close those 2. The OpenVox advisory carries separate tables for the two packages: server 9.1.0 and below and 8.17.0 and below, closed by 9.2.0 and 8.18.0, and agent 9.2.0 and below and 8.31.1 and below, closed by 9.3.0 and 8.32.0 1. 9.2.0 sits in both tables with opposite meanings, so an admin who takes the server to 9.2.0 still runs affected agents until 9.3.0 reaches them.
A low privilege requirement means a certificate here
CVSS puts privileges required at low. On this server the privilege is a certificate signed by the agent CA, and no account plays a part.
The advisory says so twice. The summary describes the attacker as holding an OpenVox or Puppet client certificate, and the details add: “In configuration with autosign (which is common for enterprise deployment), this is an unauthenticated attack chain to compromise the OpenVox / Puppet server, as a valid cert can be requested via the OpenVox / Puppet client enrollment endpoint” 1.
The score describes a server where someone has to hand you a certificate. The exploit path describes a server where you can ask for one. One setting separates the two, and it sits in puppet.conf.
Autosign, in four options
A Puppet CA signs certificate requests by hand with puppetserver ca sign, through a console, or by naive or policy-based autosigning 4.
Naive autosigning means the server signs what matches: every request, or a name pattern such as *.lab.local. Neodyme’s writeup notes that most large companies run it, because provisioning engines enroll new hosts through that endpoint 5.
Policy autosigning calls an external executable, passes it the certificate request on stdin with the certname as its argument, and takes the answer from that program. Puppet ships no default executable for this, so each installation writes its own 4. The autosign tool’s README is blunt about what most of those scripts check: a static string, and anyone who obtains the string can sign a certificate for themselves. It adds that people forget to delete csr_attributes.yaml after generating a request, and the secrets in that file spread 4.
That is why I read the setting before the score: a score assumes a configuration, and this is where the assumption sits.
The order I check
I start with the build, because it decides whether the bug sits on the host, and then the setting, because it decides who can reach it.
- The server build and the core build. The fix landed in two release lines, so I read both, and I do not accept the agent version as an answer to either.
autosignin the server section ofpuppet.conf.trueor a glob is the wide case. A path points at an executable, and then that program’s logic is the control, so I want to see what it reads before it answers.- The issued certificate list.
puppetserver ca list --allprints the certnames the CA has signed, and each one is a credential for this bug until the upgrade lands. - Port 8140. Agents reach the report endpoint on that port from wherever they run 5, and the advisory’s unauthenticated chain runs through the client enrollment endpoint 1. Confining the port to the segments that need it reduces who can ask.
- The owner of the code directory. In a hardened build the catalog belongs to a different user than
puppet, and code execution aspuppetcannot rewrite it. In a build that follows looser practice, the same code execution edits the catalog and reaches every managed host at its next check-in 5. That answer decides whether the blast radius is one server or the fleet.
What the CA key adds
Code execution as puppet “gives access to the CA signing key” 1. Neodyme’s team used that route in their lab: they copied the CA private key out, minted a certificate for the server with serverAuth and clientAuth in the extended key usage, reached a websocket broker on 8142, staged a task and read the output back from an agent 5.
The patch closes the bug on the server and leaves the question of the key open. On a host that answered on a reachable port with autosign on, I put the key in the incident scope and want certificates issued inside the exposure window reviewed. That part is my rule rather than a published one: a CA key I cannot account for is a key I replace, and its certificates inherit the doubt.
Before the upgrade lands
CISA added CVE-2026-115731 to the KEV catalog on 17 November with a 1 December deadline 6. Vox Pupuli’s guidance for the gap is to confine the server to the segments that need it and to disable or tighten autosign 7. Turning autosign off closes the enrollment path and leaves every certificate the CA issued in place, and a reader can mistake that for a fix, so I flag it when I relay the list.
Vox Pupuli also asks for a review of custom auth.conf rules after the upgrade, because some may have depended on buggy behavior fixed in the same release 7. Plain hostname rules behaved like wildcards under that behavior, so a rule that reads tight may have matched loose, and the upgrade changes that.
Open Source Puppet “will not be receiving any more updates, security or not” 7. An installation on that line needs a migration to OpenVox, since no patch is coming for it.
I keep the autosign check next to the build check on every management plane that authenticates its agents by certificate. The build tells me whether the bug is on the host, and the setting tells me whether the host hands out the credential that reaches it, so I relay both.
-
OpenVoxProject. (2026, November 3). Authenticated RCE on OpenVox server via event deserialization (GHSA-ffr6-rwpm-p5wq). GitHub. https://github.com/OpenVoxProject/openvox/security/advisories/GHSA-ffr6-rwpm-p5wq ↩︎ ↩︎ ↩︎ ↩︎ ↩︎ ↩︎
-
Perforce. (2026, November 3). CVE: Report deserialization code execution. Perforce Support Portal. https://portal.perforce.com/s/cve/a91Qi000003IaHpIAK/report-deserialization-code-execution ↩︎ ↩︎
-
National Vulnerability Database. (2026, November 3). CVE-2026-115731 detail. National Institute of Standards and Technology. https://nvd.nist.gov/vuln/detail/CVE-2026-115731 ↩︎
-
Vox Pupuli. (2026). autosign [Source code]. GitHub. https://github.com/voxpupuli/autosign ↩︎ ↩︎ ↩︎
-
Rödel, A. (n.d.). Pulling the right strings for RCE on a Puppet. Neodyme. https://neodyme.io/en/blog/puppet_rce/ ↩︎ ↩︎ ↩︎ ↩︎
-
Cybersecurity and Infrastructure Security Agency. (2026, November 17). Known exploited vulnerabilities catalog (Catalog version 2026.11.17) [Data set]. https://www.cisa.gov/known-exploited-vulnerabilities-catalog ↩︎
-
binford2k. (2026, November 3). Critical CVE updates in all known versions of Puppet and OpenVox. Vox Pupuli. https://voxpupuli.org/blog/2026/11/03/critical-cve-updates/ ↩︎ ↩︎ ↩︎