Three products treated delivery as authorization
FreeSWITCH, WordPress core and Puppet Enterprise each executed a request that arrived through a channel they were built to accept, and each fix lands in a different change-control regime.

On this page
Three products executed input that arrived through a protocol they were built to accept. FreeSWITCH read the user part of an inbound SIP MESSAGE, took the text before the plus sign as the name of a chat protocol, and ran whichever API command that name resolved to. WordPress unpacked a core release into a directory under the web root whose name followed from the archive, then served those files until the upgrade moved them. Puppet Server deserialized the body of a report submitted by any client holding a certificate its own CA had signed. None of the three needed a memory safety bug or a stolen credential.
Delivery stood in for authorization in all three. Each handler acted on a message that had passed the only check the product made. FreeSWITCH dispatched on a selector that nothing restricted. WordPress executed PHP from a staging directory that any client could request. Puppet parsed rich data types from a client whose sole qualification was a certificate the server itself had issued. The fixes land in the SIP profile, the core upgrader and the report processor, three separate changes for three separate teams.
The catalog entries arrived within a fortnight of each other, on 9, 3 and 17 November, and each records exploitation CISA treats as observed. Reach sets the order. A FreeSWITCH MESSAGE needs nothing beyond arrival, so every profile that accepts one is exposed at all times. WordPress needs the site to be inside an upgrade window, which is a race but a scheduled one. Puppet needs a certificate, which an autosigning estate hands to anyone who asks. Only the Puppet deadline, 1 December, is still open.
A SIP MESSAGE names the command the switch runs
What happened: sofia_presence_handle_sip_i_message() splits the user part of an inbound MESSAGE’s To header on a plus sign and treats the left half as a chat protocol name. The message goes to whichever chat interface registered under that name. mod_dptools registers one called api, whose address is passed to switch_api_execute(). Nothing separates a protocol that delivers text from one that runs commands, so the selector chooses the command. The SignalWire advisory states that a peer able to deliver a MESSAGE to a profile therefore chooses what the server executes.
The default configuration does not narrow that. Chat handling is enabled unless a profile sets enable-chat=false, and the stock vanilla configuration leaves it on for both the internal and external profiles. Where MESSAGE authentication is enforced, a single registered account is enough, such as an extension or a trunk peer with no rights beyond registering. The vanilla directory defines <user id="default"> with no password and no a1-hash, and such a user authenticates unless allow-empty-password is false. The inbound ACL does not cover MESSAGE, so tightening it has no effect on this path.
Execution runs as the FreeSWITCH process, which covers arbitrary OS commands, disclosure of configuration and credentials, and full call control. Where mod_lua is loaded, as the stock configuration does, inline Lua with the full standard library is a second sink. Version 1.11.3 adds a per-profile enable-chat-api-proto parameter, off unless set. A MESSAGE that selects the proto while it is disabled gets 403, logged with the source address and the requested command. The advisory states that exploitation has been observed in the wild. CISA added the flaw on 9 November with a deadline of 12 November and a forensic triage requirement.
What it reaches: FreeSWITCH sits where carrier trunks and extensions meet, so a profile that accepts MESSAGE usually speaks to the public network on 5060 or 5061. Its configuration holds the credentials that let a tenant place and receive calls, and the process that runs a command can rewrite the dialplan. On a hosted platform the process is shared across tenants, so the boundary a command crosses belongs to the platform rather than to one customer. Toll fraud and call recording follow from call control alone, before any credential is read.
What to check:
- The release running on every instance, against 1.11.3, and whether
enable-chat-api-protohas been reviewed on each SIP profile. - Whether any profile accepts inbound MESSAGE, since chat is on unless it is explicitly disabled.
- Whether the directory contains users with no password or a1-hash, including the
defaultuser from the vanilla configuration. - The user the FreeSWITCH process runs as, which caps what the command sink reaches.
- Whether logs record 403 responses to MESSAGE requests carrying a proto selector, and whether MESSAGE arrives from outside the trunk ranges.
What would change the assessment: A survey showing that most internet-facing profiles reject MESSAGE before the handler, or run with chat disabled, would narrow this to deliberate chat deployments. The advisory notes that auth-messages and disable_system_api_commands each reduce the path without closing it, so a claim that either is sufficient would not change the exposure.
The upgrade directory answers while the upgrade runs
What happened: A full-package core upgrade unpacks the downloaded release into a working directory before copying it over the installed tree. In 7.1.3 and earlier, WP_Upgrader::unpack_package() built that name from the package file name with the .tmp and .zip suffixes removed, and placed it under wp-content/upgrade/, which the 7.1.4 source describes as web-served. The staged tree is a complete WordPress installation, including wp-admin/setup-config.php, which runs without authentication and offers to create wp-config.php.
A request for the staged copy of that script during the window writes a configuration file into the working directory. The upgrade then copied everything from that directory into the installed tree except wp-content and wp-includes/version.php, so a planted wp-config.php replaced the real one. The 7.1.4 change adds wp-config.php to the skip list, appends a sixteen-character random suffix from wp_generate_password() to the working directory name, and writes an index.php into wp-content/upgrade/ to prevent listing.
The weakness is a race, which is what the 8.1 score with high attack complexity records. CISA added it on 3 November with a deadline of 6 November and a forensic triage requirement, which places observed exploitation inside upgrade windows rather than against idle sites. The upgrade that installs 7.1.4 runs the 7.1.3 upgrader, so the last vulnerable window is often the one that closes it.
What it reaches: The affected range runs from 4.7.0 to 7.1.3, every release since 2016. A site is exposed only while a full-package upgrade runs, and that window is short on one host. Automatic background updates create it on a schedule an attacker can wait for. A fleet spreads those windows across thousands of hosts, so the population exposed at any hour is large even though each window is brief. wp-config.php holds the database credentials and authentication salts. Replacing it points the installation at a database the attacker controls, and administrator access and code execution follow through the ordinary editing paths.
What to check:
- The core version reported by every site in the estate, against 7.1.4.
- Whether
wp-content/upgrade/exists, what it contains, and whether the web server serves it from the internet, tested by request rather than by inspection. - Access logs for requests to PHP files under
wp-content/upgrade/, particularlywp-admin/setup-config.php. - Whether
wp-config.phpchanged during a core upgrade since 2026-10-28, compared by modification time and hash against the deployment record.
What would change the assessment: Evidence that the exposed population disables background updates, or blocks wp-content/upgrade/ at the edge, would cut the window to manual upgrades. The opposite finding, that a staged tree survives a failed upgrade, would remove the timing condition and raise the score. Nothing in the reviewed sources establishes that persistence.
A report from any signed agent certificate
What happened: Puppet Server collects reports from agents over HTTPS and deserializes the report body, including the rich data types that carry structured values. A client holding any certificate the CA has signed can submit a crafted rich data type in a PUT request, and the server deserializes it and executes code as the service user. The CNA record assigns CWE-502 and CWE-94. NIST scores the flaw 9.9 under CVSS 3.1 with a scope change, Perforce 9.4 under CVSS 4.0.
Where the certificate authority signs requests automatically without strong verification, an unauthenticated network attacker can obtain a valid certificate and submit the report. Autosigning makes enrolling a node a single step, and estates that use it for convenience extend that to any host that can reach port 8140. Perforce fixed the flaw in Puppet Enterprise 2023.8.13, 2025.12.1 and 2026.0.1. The CNA record lists affected ranges only for Enterprise, and the reviewed sources do not establish open-source Puppet Server’s exposure.
CISA added the flaw on 17 November with a deadline of 1 December, and the catalog records that exploitation has been observed. EPSS sits at 0.007, in the 54th percentile. That is the lowest of the three and the least informative, because the score models attempts across the internet rather than estates with autosigning switched on.
What it reaches: Puppet Server compiles the catalog every managed node applies, holds the CA that issues agent certificates, and receives the reports that describe what each node did. Code running as that service can read catalogs, which routinely carry the credentials and secrets manifests distribute. It can also alter the classes agents fetch, so the next run on every node applies whatever the server returns. It can also mint certificates, which turns one agent identity into control of the CA and outlives the patch on any node enrolled while the flaw was open.
What to check:
- The Puppet Enterprise version against 2023.8.13, 2025.12.1 and 2026.0.1.
- Whether the CA signs automatically, through
autosign.confentries, a policy-based autosign executable, or a wildcard entry. - Whether port 8140 answers anything outside the managed subnets, which is the population that can present a certificate at all.
- Which certificate names exist in the CA without a corresponding managed node, particularly those issued since 2026-11-03.
What would change the assessment: A policy-based autosign configuration that ties signing to pre-registered hostnames would make a certificate hard to obtain, which separates a fleet-wide flaw from an authenticated one. A public exploit would move the other way. The reviewed sources describe neither, and the catalog entry records exploitation without saying what the initial certificate was.
The three fixes land in change-control regimes that move at different speeds. A SIP profile change is a configuration file on a telephony host, and it needs a restart of the switch. A core upgrade needs a maintenance window on every site, which is why the WordPress window recurs on a schedule. A Puppet Server patch changes the version on the single host that holds the CA, and the CA has to keep signing through the change. Two of the three deadlines, 6 and 12 November, have passed.
