PostgreSQL released updates yesterday for 52 security flaws, and the one that will cause the most operational pain is not the most severe one on the list.
CVE-2026-6460 is a type confusion flaw in composite types. It scores 8.8, it lets a database user who can create objects run code on the server as the operating system user running PostgreSQL, and fixing it properly requires rebuilding every C-language extension you have installed.
That last requirement is where deployments will go wrong.
What the Flaw Actually Does
Composite types in PostgreSQL are built from named fields, each with its own data type. The server holds values of those types in various places across its memory.
A field's type can be changed while older values of that type are still held elsewhere. The server can then read an old value as if it carried the new type, and that mismatch is the confusion.
The fix, described in the release notes, adds version checks so the server can detect a value that no longer matches its type.
The Root Cause Is Twenty Years Old
Mehmet Ince traced the flaw's origin to May 5, 2004, when PostgreSQL gained the ability to change a column's type. His technique for turning it into code execution relies on a later change made on May 14, 2015.
That timeline is worth noting, because it means the underlying condition has existed through two decades of releases.
Ince published a working exploit demonstrated on PostgreSQL 18.6 this week. The writeup includes the SQL that crashes an unpatched server and a second step that leaks memory addresses. It stops short of the final step that makes the server execute a command.
The Privilege Question
Ince wrote that he could trigger the crash as a lowest-privileged login role. The project's advisory describes the attacker only as someone who can create objects, without specifying which privileges that takes. Red Hat's CVE assessment says the attacker needs a database login with the ability to create or modify objects.
None of those sources answers the practical question. Whether an account with only default privileges can go beyond the crash into code execution has not been stated.
That gap matters if you are deciding how fast to patch a database with many application roles.
The Extension Rebuild Requirement
Patching the server is necessary and not sufficient.
Every C-language extension built against an older release has to be rebuilt or updated at the same time. The RPM package maintainers put it directly in a repository notice. Extensions built against older releases still load, but they may not be protected from the vulnerabilities fixed in this release.
Three complications follow from that.
- Some extensions need source changes. Not just a rebuild. The code itself requires modification.
- Rebuilt extensions can fail on older servers. If you run a mixed-version fleet, the new extension package may not load on the old minor releases.
- The RPM repositories now enforce it. Extension packages in the project's RPM repositories require the new minor release, which removes the option of staggering the rollout quietly.
The Managed Service Coordination
Ince adapted the exploit for managed PostgreSQL services and privately shared proofs of concept with their security teams in August.
The providers he worked with include AWS (Aurora and RDS), Microsoft (Azure PostgreSQL), Google (AlloyDB), Supabase, Databricks, and Ubicloud.
The writeup does not give each service's current patch status. If your database runs on one of them, ask directly rather than assuming.
The Rest of the 52 Flaws
Thirty-two of the 52 flaws are rated 8.8, which is the highest score in the release. That is a lot of fixes at the top of the severity range.
Based on the scores in the project's vulnerability index, only one flaw in the server requires neither a login nor any user action. That one is CVE-2026-116322, which lets an unauthenticated client trigger a restart of PostgreSQL 18 servers.
Several fixes land in client tools, and those matter because they often run with more privilege than people remember.
|
Tool |
Exposure |
|
pg_restore |
A crafted archive can make it run code as the user running it |
|
psql via pg_dump |
A compromised source server can make psql run code through the restore script in pg_dump's tar output |
Affected Versions and Fixes
|
Branch |
Fixed Release |
Status |
|
18 |
18.7 |
Supported |
|
17 |
17.12 |
Supported |
|
16 |
16.16 |
Supported |
|
15 |
15.20 |
Supported |
|
14 |
14.25 |
End of life, final release |
|
13 and older |
None |
No update |
PostgreSQL 14 reached end of life, and 14.25 is its last release. Older major versions are affected and received no update at all.
Exploitation Status
The PostgreSQL Global Development Group has not reported any attacks using the flaw. CISA's assessment in the CVE record also lists exploitation as none.
The project credits more than a dozen reporters for CVE-2026-6460, including Ince, and credits him on five other flaws in the same release.
What Administrators Should Do
- Update the server to 18.7, 17.12, 16.16, 15.20, or 14.25, matching your major version.
- Rebuild or update every C-language extension at the same time. Not after. The window between patching the server and patching extensions is a window where the server is protected and the extensions are not.
- Check whether your extensions need source changes. A rebuild alone does not cover all of them.
- Sequence the rollout if you run mixed versions. Rebuilt extensions can fail to load on older servers, and the RPM packages now require the new minor release.
- Restart immediately if PostgreSQL runs outside its package's service unit. Under Patroni, for example. A maintainer of the RPM packages says so in a packaging blog post.
- Update client tools everywhere they run. pg_dump, pg_restore, reindexdb, vacuumdb. These are frequently installed on application servers and workstations that nobody thinks of as database hosts.
- Do not disable the new composite_type_mismatch setting. The project says turning off the checks carries security risks.
- Follow the reindexing steps in the release announcement.
- Take a new full backup if you use incremental backups, and base later incrementals on it.
The advisory and release notes offer no workaround for servers that cannot be updated yet.
FAQ
What is CVE-2026-6460?
A type confusion flaw in PostgreSQL's composite types. A field's type can be changed while older values of that type are still held elsewhere, and the server can then read an old value as if it had the new type. It lets a database user who can create objects run code on the server.
What is the CVSS score?
8.8. It is one of 32 flaws in the release rated at that level.
Which versions are affected?
All versions before 18.7, 17.12, 16.16, 15.20, and 14.25.
Why do I need to rebuild extensions?
Because extensions built against older releases still load, but they may not be protected from the vulnerabilities fixed in this release. The RPM package maintainers state this directly.
Will every extension rebuild cleanly?
No. Some extensions need source-code changes, not just a rebuild. And extensions rebuilt for the new releases can fail to load on older servers.
How old is this flaw?
The underlying condition traces to May 5, 2004, when PostgreSQL gained the ability to change a column's type. The code execution technique relies on a change from May 14, 2015.
Has it been exploited?
The PostgreSQL Global Development Group has not reported attacks, and CISA's assessment lists exploitation as none.
Can a lowest-privileged account exploit it?
Ince says he could trigger the crash as a lowest-privileged login role. The project describes the attacker as someone who can create objects. Red Hat says the attacker needs a login with the ability to create or modify objects. None of those sources says whether default privileges alone can reach code execution.
Is there an unauthenticated flaw in this release?
Yes. CVE-2026-116322 lets an unauthenticated client trigger a restart of PostgreSQL 18 servers. It is the only server flaw in the release that requires neither a login nor user action.
Are the client tools affected?
Yes. A crafted archive can make pg_restore run code as the user running it, and a compromised source server can make psql run code through the restore script in pg_dump's tar output.
Is PostgreSQL 14 still supported?
No. PostgreSQL 14 reached end of life, and 14.25 is its final release.
What about managed services?
Ince adapted the exploit for AWS Aurora and RDS, Azure PostgreSQL, Google AlloyDB, Supabase, Databricks, and Ubicloud, and shared proofs of concept with their security teams in August. Their current patch status is not stated in the writeup.
Is there a workaround?
No. The advisory and release notes offer none.