Any PostgreSQL role that can create a table on your server is the one to worry about this week. CVE-2026-6460 is a type confusion bug in how the server handles composite values after DDL. The PostgreSQL project rates it CVSS 8.8 because an object creator can turn it into arbitrary code running as the operating system user behind the database, and on a cluster first built on PostgreSQL 14 or older that can be every login role you have.
We ran every command below in November 2026 on four fresh GCP VMs: Ubuntu 24.04 with PGDG PostgreSQL 14 and 17 side by side, Debian 12 and 13 on distro packages, and Rocky Linux 9 with PGDG PostgreSQL 18 and pgvector. That spread matters, because the fix reached each package source on a different day and in a different shape, and one of them still has no fixed build.
You’ll get a check for each cluster, an audit of which roles count as attackers, the patch path per distro, and what to do on hosts you cannot patch tonight.
The fixed PostgreSQL releases for CVE-2026-6460
Upstream shipped the fix on 12 November 2026 in the minor releases below, and every earlier minor of every supported branch is affected. The official CVE-2026-6460 advisory names ALTER ... SET DATA TYPE as the main route and lists other DDL routes as well, so there is no single statement you can block instead of patching. Treat anything below these numbers as exposed.
| Branch | Fixed in | server_version_num at or above | Notes |
|---|---|---|---|
| PostgreSQL 18 | 18.7 | 180007 | Current stable branch |
| PostgreSQL 17 | 17.12 | 170012 | Debian 13 default |
| PostgreSQL 16 | 16.16 | 160016 | Ubuntu 24.04 default |
| PostgreSQL 15 | 15.20 | 150020 | Debian 12 default |
| PostgreSQL 14 | 14.25 | 140025 | Last minor, branch is now end of life |
| PostgreSQL 13 and older | No fix | None | Already end of life, upgrade the major version |
The same release fixes 52 CVEs in total, according to the PostgreSQL release announcement, and our roll-up of the November PostgreSQL security update walks through them one by one. We single this one out for two reasons: it is the one that reaches code execution, and it is the one whose fix changes how C extensions have to be built. Both shape the patch plan below, so read the extension section before you schedule anything.
A version check for every cluster on a host
Package versions lie when a cluster has not restarted since the upgrade, so ask the running server instead of dpkg or rpm. On Debian and Ubuntu, pg_lsclusters lists every cluster the host runs, which matters on machines that carry two majors side by side like our Ubuntu VM did:
pg_lsclusters
Two clusters came back, PostgreSQL 14 on the default port and 17 on the next one up.
Ver Cluster Port Status Owner Data directory Log file
14 main 5432 online postgres /var/lib/postgresql/14/main /var/log/postgresql/postgresql-14-main.log
17 main 5433 online postgres /var/lib/postgresql/17/main /var/log/postgresql/postgresql-17-main.log
Rather than connect to each port by hand, loop over the list and ask every running server for its version string and numeric version:
for port in $(pg_lsclusters -h | awk "{print \$3}"); do sudo -u postgres psql -p $port -Atc "SELECT current_setting('port'), current_setting('server_version'), current_setting('server_version_num')"; done
Both clusters turned out to be one minor short of the fix.
5432|14.24 (Ubuntu 14.24-1.pgdg24.04+2)|140024
5433|17.11 (Ubuntu 17.11-1.pgdg24.04+2)|170011
Compare the third column with the table above: 140024 is below 140025 and 170011 is below 170012, so both are vulnerable. We’d feed the numeric form into monitoring rather than the version string, because a plain integer comparison works across every branch and does not trip on distro suffixes like pgdg24.04+2. One threshold per major is all the alert needs.

RHEL family hosts with PGDG packages have no pg_lsclusters, so connect to each instance’s port directly. Those builds report the bare number (18.6 on our Rocky VM) with no distro tag, and the same numeric thresholds apply.
The composite_type_mismatch parameter as a fix marker
Distros sometimes backport fixes without bumping the upstream number, so a version string alone can mislead you in either direction. The fixed releases add a new server parameter, composite_type_mismatch, that unpatched builds do not have. Passing true as the second argument to current_setting returns an empty value instead of an error when the parameter is missing:
sudo -u postgres psql -Atc "SELECT current_setting('server_version'), current_setting('composite_type_mismatch', true)"
On the unpatched Debian 12 VM the second field came back blank.
15.18 (Debian 15.18-0+deb12u1)|
A blank means the running binary predates the fix, while a patched server prints error there, the default. We lean on that marker after each upgrade below and would wire it into fleet inventory too, because it answers “does this binary carry the fix” instead of “what did the packager call it”. On a host where the two disagree, believe the marker.
Distro and repository status on 14 November 2026
Two days after the upstream release, the distros are moving at very different speeds. Here is where each source stood when we patched the lab. Recheck your own distro’s tracker before you rely on any “no fix yet” row, because those rows change daily.
| Source | Package state | What we saw in the lab |
|---|---|---|
| PGDG apt (apt.postgresql.org) | Fixed for 14 to 18 | 14.25 and 17.12 installed, marker reads error |
| PGDG yum (download.postgresql.org) | Fixed for 14 to 18 | 18.7 installed on Rocky Linux 9, marker reads error |
| Debian 13 trixie | Fixed, DSA-6625-1 | 17.12-0+deb13u1 from stable-security |
| Debian 12 bookworm | 15.20-0+deb12u1 in oldstable-security, tracker still lists it as vulnerable, no DSA yet | The 15.20 build carries the marker, so the fix code is in it |
| Ubuntu 26.04, 24.04, 22.04 | Needs evaluation, no USN | Archive still serves 16.15 on 24.04 |
| RHEL, Rocky Linux, AlmaLinux AppStream | No advisory for RHEL 8, 9 or 10 | Red Hat lists only its Hardened Images as affected so far |
Ubuntu is the row that catches people. Ubuntu 24.04 ships PostgreSQL 16 in main, and on our VM the archive’s newest build was still the unpatched one while PGDG already had the fix. You can see both side by side in the package policy:
apt-cache policy postgresql-16 | head -12
The PGDG candidate sits at the top, the Ubuntu security pocket one minor behind it.
postgresql-16:
Installed: (none)
Candidate: 16.16-1.pgdg24.04+1
Version table:
16.16-1.pgdg24.04+1 500
500 https://apt.postgresql.org/pub/repos/apt noble-pgdg/main amd64 Packages
16.15-1.pgdg24.04+2 500
500 https://apt.postgresql.org/pub/repos/apt noble-pgdg/main amd64 Packages
16.15-0ubuntu0.24.04.1 500
500 http://europe-west1.gce.archive.ubuntu.com/ubuntu noble-updates/main amd64 Packages
500 http://security.ubuntu.com/ubuntu noble-security/main amd64 Packages
16.14-1.pgdg24.04+1 500
If you run Ubuntu’s own postgresql-16 package, apt upgrade has nothing to give you yet. Your choices are to wait for a USN with the mitigations further down in place, or to move the host to the PGDG repository. Our guide on installing PostgreSQL 18 on Ubuntu 26.04 sets that repository up with the apt.postgresql.org.sh script we used on the lab VM.

Red Hat rates the CVE Important and, for RHEL itself, offers only mitigations so far: firewall the port and revoke CREATE on the public schema. Rocky Linux and AlmaLinux AppStream follow RHEL, so hosts on the stock modules wait with it unless they switch to PGDG.
Who counts as an object creator on your cluster
The advisory’s attacker is “an object creator”, so your exposure depends on which login roles can create anything. Superusers already own the box and don’t change the picture. Ordinary application and reporting logins are what matter, and how much they can create depends on how the cluster was born:
- Clusters created on PostgreSQL 15 or newer: PUBLIC lost CREATE on the
publicschema in 15, so a plain login role cannot create tables unless someone granted it. - Clusters created on 14 or older: every login role can create in
publicin every database, and this stays true after a dump and restore or apg_upgradeonto a newer major. - Explicit grants: schema owners, roles with CREATE on an app schema, and database owners all count, whatever the version.
- Temporary objects: PUBLIC holds TEMPORARY on every database you create (
template1excepted), so any login can create temp tables there. Treat that as a possible route too until the fix is in.
The second point is the trap. A cluster moved forward through a PostgreSQL major version upgrade keeps the old grant, so “we run 17” says nothing about whether your app login can create tables. Ask the catalog instead of trusting the version.
An audit query for CREATE and TEMP rights
Save the two queries below to a file so you can run them against every database. The first lists non-superuser logins that can create in any user schema, and the second shows which databases each login can add schemas to or create temp tables in.
sudo vim /var/tmp/audit-create.sql
Paste in both queries:
SELECT r.rolname AS role, n.nspname AS schema
FROM pg_roles r CROSS JOIN pg_namespace n
WHERE r.rolcanlogin AND NOT r.rolsuper
AND n.nspname NOT LIKE 'pg\_%' AND n.nspname <> 'information_schema'
AND has_schema_privilege(r.oid, n.oid, 'CREATE')
ORDER BY 1, 2;
SELECT r.rolname AS role, d.datname AS database,
has_database_privilege(r.oid, d.oid, 'CREATE') AS create_schema,
has_database_privilege(r.oid, d.oid, 'TEMP') AS temp_tables
FROM pg_roles r CROSS JOIN pg_database d
WHERE r.rolcanlogin AND NOT r.rolsuper AND d.datallowconn
ORDER BY 1, 2;
Schema privileges live per database, so run the file in each one. On our Ubuntu host both clusters had the same app_user login and an appdb database, and the PostgreSQL 14 cluster on port 5432 went first:
sudo -u postgres psql -p 5432 -d appdb -f /var/tmp/audit-create.sql
Nobody granted app_user anything, yet the plain application login can create tables in public:
role | schema
----------+--------
app_user | public
(1 row)
role | database | create_schema | temp_tables
----------+-----------+---------------+-------------
app_user | appdb | f | t
app_user | postgres | f | t
app_user | template1 | f | f
(3 rows)
Running the same file against the PostgreSQL 17 cluster on port 5433 returned zero rows in the first result and the identical TEMP table in the second. On 17, then, the login cannot create a real table without a grant, yet it can still create temp tables in appdb and postgres. All of it came from defaults, which is why we run this audit on every cluster rather than guessing from the major version.

Any row in the first result moves a host from “patch this month” to “patch tonight”. Shared hosting, SaaS tenants with their own logins, BI users with SQL access, and anything where a web app’s login could be hijacked through SQL injection all land there, because each of those hands DDL to someone you do not fully control.
Extensions need the new minor too
Unlike most minor releases, this fix reaches past the server. The PostgreSQL release notes say full protection requires rebuilding C-language extensions, and the PGDG notes explain why: the fix turns DatumGetHeapTupleHeader() from a macro into a server function and adds a new check inside heap_getattr(), both of which an extension only picks up when it is compiled again. PGDG rebuilt its extension packages and tied each one to the fixed minor, which shows up in the pgvector dependencies:
apt-cache show postgresql-17-pgvector=0.8.7-1.pgdg24.04+2 | grep -E "^(Version|Depends)"; apt-cache show postgresql-17-pgvector=0.8.7-1.pgdg24.04+1 | grep -E "^(Version|Depends)"
The rebuild (+2) refuses to install next to anything older than the fixed server, where the old build accepted any 17:
Version: 0.8.7-1.pgdg24.04+2
Depends: postgresql-17 (>> 17.12), libc6 (>= 2.38)
Version: 0.8.7-1.pgdg24.04+1
Depends: postgresql-17, libc6 (>= 2.38)
RPMs do the same thing through the libs package. On Rocky Linux 9 the rebuilt pgvector_18 requires postgresql18-libs(x86-64) >= 18.7, and the PGDG yum notice spells out what that means in practice: updating any extension now drags the server to the new minor and restarts it.

Two practical rules follow from those dependencies:
- One transaction: upgrade the server and its extensions together, never an extension alone on a host you meant to patch later, or the server moves under you at a time you did not pick.
- Out-of-repo builds: list extensions that came from a vendor
.soor amake install, because nobody rebuilt those for you. They keep loading, but they may not be protected.
The pgvector setup in our vector database comparison is the typical case: an extension most teams install from a repo once and then forget it exists.
Patching PGDG clusters on Ubuntu and Debian
With the PGDG apt repository in place, the fixed packages show up after a normal index refresh. Our Ubuntu 24.04 host listed both majors, the rebuilt pgvector and a newer postgresql-common:
sudo apt-get update -qq && apt list --upgradable 2>/dev/null | grep -i postgres
Seven packages were waiting, including both common packages.
postgresql-14/noble-pgdg 14.25-1.pgdg24.04+1 amd64 [upgradable from: 14.24-1.pgdg24.04+2]
postgresql-17-pgvector/noble-pgdg 0.8.7-1.pgdg24.04+2 amd64 [upgradable from: 0.8.7-1.pgdg24.04+1]
postgresql-17/noble-pgdg 17.12-1.pgdg24.04+1 amd64 [upgradable from: 17.11-1.pgdg24.04+2]
postgresql-client-14/noble-pgdg 14.25-1.pgdg24.04+1 amd64 [upgradable from: 14.24-1.pgdg24.04+2]
postgresql-client-17/noble-pgdg 17.12-1.pgdg24.04+1 amd64 [upgradable from: 17.11-1.pgdg24.04+2]
postgresql-client-common/noble-pgdg 295.pgdg24.04+1 all [upgradable from: 257build1.1]
postgresql-common/noble-pgdg 295.pgdg24.04+1 all [upgradable from: 257build1.1]
Upgrade only the PostgreSQL packages so the maintenance window does not also pull in a kernel. Name every server, client and extension package installed for each major on the host:
sudo DEBIAN_FRONTEND=noninteractive apt-get install -y --only-upgrade postgresql-14 postgresql-17 postgresql-client-14 postgresql-client-17 postgresql-17-pgvector postgresql-common postgresql-client-common
Here is the relevant part of the apt output.
The following packages will be upgraded:
postgresql-14 postgresql-17 postgresql-17-pgvector postgresql-client-14
postgresql-client-17 postgresql-client-common postgresql-common
7 upgraded, 0 newly installed, 0 to remove and 11 not upgraded.
...
Setting up postgresql-14 (14.25-1.pgdg24.04+1) ...
Setting up postgresql-17 (17.12-1.pgdg24.04+1) ...
Setting up postgresql-17-pgvector (0.8.7-1.pgdg24.04+2) ...
No services need to be restarted.
No containers need to be restarted.
Don’t read “No services need to be restarted” as “nothing restarted”. That line comes from needrestart running after the fact, once the package scripts had already stopped and started both clusters. The next check proves it, and it matters for planning because every client connection on the host drops at that moment.

Rerun the version loop, this time with pg_postmaster_start_time() in place of the numeric version so the restart is visible:
for port in $(pg_lsclusters -h | awk "{print \$3}"); do sudo -u postgres psql -p $port -Atc "SELECT current_setting('port'), current_setting('server_version'), pg_postmaster_start_time()"; done
Both clusters now run the fixed minors.
5432|14.25 (Ubuntu 14.25-1.pgdg24.04+1)|2026-11-14 01:33:47.182602+00
5433|17.12 (Ubuntu 17.12-1.pgdg24.04+1)|2026-11-14 01:33:47.187653+00
Both start times read 01:33:47, the second the package scripts finished, so every open connection on the host dropped during the install. On a lab VM nobody notices; on a primary it is a real outage, which is why the fleet order at the end of this article puts replicas first. Connection poolers such as PgBouncer hide some of it, since they reconnect on their own, but any transaction in flight at that second still fails and your application has to retry it.
Patching Debian’s own PostgreSQL packages
Hosts on Debian’s packages need nothing beyond the security archive, which our PostgreSQL on Debian guide leaves enabled by default. On Debian 12 the fixed build sits in oldstable-security:
apt list --upgradable 2>/dev/null | grep -i postgres
Only the server and its client were pending.
postgresql-15/oldstable-security 15.20-0+deb12u1 amd64 [upgradable from: 15.18-0+deb12u1]
postgresql-client-15/oldstable-security 15.20-0+deb12u1 amd64 [upgradable from: 15.18-0+deb12u1]
Install those two with the same targeted upgrade:
sudo DEBIAN_FRONTEND=noninteractive apt-get install -y --only-upgrade postgresql-15 postgresql-client-15
Then rerun the marker check from earlier.
sudo -u postgres psql -Atc "SELECT current_setting('server_version'), current_setting('composite_type_mismatch', true)"
The parameter that was blank before now reads error:
15.20 (Debian 15.20-0+deb12u1)|error
Here is why we don’t trust a tracker alone. Debian’s security tracker still listed bookworm as vulnerable on 14 November, yet the build it serves carries the fix code. Our reading is that the advisory is pending rather than the package being incomplete, but recheck the tracker for a DSA before you close the ticket. Auditors tend to ask for the advisory number, and on 14 November there wasn’t one to give them for bookworm, so record the marker output alongside the package version as your evidence.

Debian 13 runs the same sequence on postgresql-17 from stable-security, covered by DSA-6625-1. Our trixie VM went from a blank marker to this:
17.12 (Debian 17.12-0+deb13u1)|error
One gap remains on trixie: Debian’s postgresql-17-pgvector was still the original build from the main archive after the server update.
postgresql-17-pgvector:
Installed: 0.8.0-1
Candidate: 0.8.0-1
Version table:
*** 0.8.0-1 500
500 mirror+file:/etc/apt/mirrors/debian.list trixie/main amd64 Packages
Nothing breaks: the extension still loads and answered a distance query against the patched server. It just hasn’t been rebuilt against the new headers, which is exactly the “works but may not be protected” state the release notes warn about. If you run Debian’s extension packages, watch for rebuilt uploads over the coming weeks.

Teams that need a protected pgvector on Debian today can take the rebuilt extension from PGDG instead, but we would move the server to PGDG packages at the same time rather than mix a PGDG extension with a Debian server build. Mixed sources are how a later update leaves the two out of step again.
Patching PGDG PostgreSQL on Rocky Linux, AlmaLinux and RHEL
Since AppStream has no fixed build yet, the PGDG yum repository is the only patched source on the RHEL family today. Our Rocky Linux 9 VM ran PGDG PostgreSQL 18 with pgvector, installed the usual way with the AppStream module disabled, and started out like this:
rpm -q postgresql18-server pgvector_18; sudo -u postgres psql -Atc "SELECT current_setting('server_version'), pg_postmaster_start_time()"
The server was on the last unpatched minor.
postgresql18-server-18.6-1PGDG.rhel9.8.x86_64
pgvector_18-0.8.7-1PGDG.rhel9.8.x86_64
18.6|2026-11-14 01:57:57.910783+00
Upgrade the server packages and the extension together. The glob covers the server, libs and client for one major, and every other extension package you run belongs on the same line:
sudo dnf upgrade -y "postgresql18*" pgvector_18
dnf pulled the libs first and the extension last, as the dependency demands:
Upgrading:
pgvector_18 x86_64 0.8.7-2PGDG.rhel9.8 pgdg18 115 k
postgresql18 x86_64 18.7-1PGDG.rhel9.8 pgdg18 2.0 M
postgresql18-libs x86_64 18.7-1PGDG.rhel9.8 pgdg18 314 k
postgresql18-server x86_64 18.7-1PGDG.rhel9.8 pgdg18 7.3 M
...
Upgraded:
pgvector_18-0.8.7-2PGDG.rhel9.8.x86_64
postgresql18-18.7-1PGDG.rhel9.8.x86_64
postgresql18-libs-18.7-1PGDG.rhel9.8.x86_64
postgresql18-server-18.7-1PGDG.rhel9.8.x86_64
Complete!
You don’t need to restart postgresql-18 by hand, because the PGDG server package restarts a running service on a minor update, as the PGDG notice linked above says. Rerunning the rpm -q line confirms it:
postgresql18-server-18.7-1PGDG.rhel9.8.x86_64
pgvector_18-0.8.7-2PGDG.rhel9.8.x86_64
18.7|2026-11-14 01:58:19.824552+00
The postmaster start time moved from 01:57:57 to 01:58:19, the moment dnf finished, so the RPM restarted the service just as the Debian packages did. Plan the window the same way on both families, and don’t let a config management run that applies all pending updates loose on a primary during business hours. A SHOW composite_type_mismatch afterwards returned error, and a pgvector distance query still worked against the upgraded extension.

If you’re on the stock AppStream modules and cannot switch repositories, skip ahead to the mitigation section and keep watching your distro’s errata page.
Leave composite_type_mismatch at error
Patched servers accept three values for the new parameter, which you can read straight from the catalog:
sudo -u postgres psql -p 5433 -Atc "SELECT enumvals FROM pg_settings WHERE name = 'composite_type_mismatch'"
Our patched 17 cluster listed three.
{error,warn,disabled}
Each value trades protection against compatibility differently:
- error: the default and the only setting that protects you; a mismatched composite value stops the query.
- warn: logs the mismatch and carries on, which leaves the bug reachable.
- disabled: skips the check entirely, and the release notes say outright that this carries security risks.
The last two exist for an extension that breaks after the upgrade because it builds composite values the old way. The parameter’s context is superuser, so only superusers or roles granted SET on it can flip it, which is exactly the admin who might do so to quiet a failing extension. If one fails, update or rebuild it instead of loosening the check.
Mitigation for hosts you can’t patch yet
Ubuntu archive users, AppStream users and anyone waiting on a change window can still take away the attacker’s foothold. None of this replaces the patch, but each step shrinks who can reach the bug:
- Revoke CREATE on public: run
REVOKE CREATE ON SCHEMA public FROM PUBLICin every database of clusters born on 14 or older, plustemplate1so new databases inherit it. - Trim explicit grants: remove CREATE on app schemas from roles that only read and write data, and give migrations their own role.
- Close the network path: restrict
pg_hba.confand the host firewall to the application hosts that need the port. - Watch for odd DDL:
log_statement = 'ddl'recordsALTER TABLEfrom roles that never alter tables.
On the PostgreSQL 14 cluster, the revoke followed by the audit looks like this:
sudo -u postgres psql -p 5432 -d appdb -c "REVOKE CREATE ON SCHEMA public FROM PUBLIC" && sudo -u postgres psql -p 5432 -d appdb -f /var/tmp/audit-create.sql
The schema list is now empty, but the TEMP column has not moved:
REVOKE
role | schema
------+--------
(0 rows)
role | database | create_schema | temp_tables
----------+-----------+---------------+-------------
app_user | appdb | f | t
app_user | postgres | f | t
app_user | template1 | f | f
(3 rows)
That remaining t is the limit of the mitigation. Temp tables are still objects the role creates, and revoking TEMP from PUBLIC breaks plenty of reporting and ORM code, so we would only revoke it from PUBLIC on a database where we can grant TEMPORARY back to the roles that need it. Red Hat also warns that the public schema revoke can break apps that create tables at runtime, so try it on staging first.

Keep the revoke after you patch. PostgreSQL has shipped the same default since 15, and it closes the door for whatever DDL bug comes next.
PostgreSQL 14 hosts after this release
PostgreSQL 14 will never get another minor after this one. Patch it now, because the fix is real, but the next CVE in this class won’t have a 14 build at all, and PostgreSQL 13 or older has none today. Those old clusters are also the ones most likely to hand CREATE on public to every login, as our audit showed. Plan the move with our pg_upgrade walkthrough, and take a fresh base backup first using our backup and point-in-time recovery guide.
Rolling the patch across a replicated fleet
Every package source we tested restarted PostgreSQL during the install, so the order of hosts decides how much downtime you take. Minor releases are binary compatible within a major, which means streaming replicas on the new minor follow a primary on the old one without trouble. That gives the usual sequence:
- Inventory first: run the version and marker check on every instance and record which still return a blank marker.
- Replicas next: patch each standby, confirm it reconnects and that lag returns to normal before the next one.
- Switch over: promote a patched replica, then patch the old primary and rejoin it as a standby.
- Audit extension sources: list extensions per database with
\dx, then check that each library under the per-majorpg_config --pkglibdir(/usr/lib/postgresql/17/bin/pg_configon Debian and Ubuntu,/usr/pgsql-18/bin/pg_configon PGDG RPM hosts) belongs to a package withdpkg -Sorrpm -qf. A file no package owns came from a manual build and needs a rebuild. - Follow the migration notes: the release notes for each branch list index corruption fixes that need a REINDEX of affected indexes, and the 17.12 notes add that existing incremental backups should not be trusted, so take a new full backup after updating if you use them.
On a Patroni cluster the switchover step is a patronictl switchover rather than a manual promote, and our Patroni, etcd and HAProxy setup shows how HAProxy moves clients across without a config change. On managed services such as RDS, Azure Database for PostgreSQL and AlloyDB the provider applies the minor release, so follow its maintenance notices there instead of this procedure.