The PostgreSQL developers released versions 18.7, 17.12, 16.16, 15.20 and 14.25 on November 12, 2026, alongside the second release candidate of PostgreSQL 19. What the project's calendar calls a minor update has shipped as the largest security package in its recent history: 52 vulnerabilities fixed and more than 140 bugs, according to the official PostgreSQL vulnerability registry.
The figure that sets the urgency is not the volume but the score. Thirty-two of the 52 CVEs share a CVSS 3.1 rating of 8.8, the highest in the batch, and almost all of them describe the same outcome: type confusion or freed memory that ends in code execution with the privileges of the operating system user running the database. The lowest in the package is a 3.1 that allows bypassing the EXECUTE privilege on conversion functions.
There is a second piece of news inside the announcement. Version 14.25 is the last in the 14 branch, which from now on receives no security patches or bug fixes. Anyone still on that branch has a calendar problem rather than a configuration one: every month without migrating is a month without a safety net, and jumping to 15 or 16 requires a maintenance window that nobody reserves for fun.
Five versions released on the same day
The batch arrived in five versions at once: 18.7 and 17.12; 16.16 and 15.20; and 14.25. The update is cumulative, so there is no need to go through pg_upgrade or export the data: simply stop the server, replace the binaries and start it again.
It is worth checking which of those five branches is running on each machine before deciding anything, because the oldest supported branch is now 15 and the announcement closes the door on 14 without ceremony. The PostgreSQL versioning policy sets five years of support per branch from its first release.
Almost all the remote code execution flaws share the same technical shape. PostgreSQL handles values with a type tag and, when that tag and the actual content stop matching, the engine ends up interpreting arbitrary bytes as if they were another structure. The result ranges from a memory read to control of the execution flow.
Code and legacy language
The names in the list map out the code's hot spots. PL/Perl and PL/Tcl are in the embedded language layer; ts_rank_cd() in full-text search; parallel query and logical replication inside the engine; and dblink with statistics import at the edge of the extensions. These are pieces that are rarely audited with the same scrutiny as the core and that are, at the same time, exposed to any user with connection permission.
| CVE |
CVSS 3.1 |
What it allows |
| CVE-2026-6460 |
8.8 |
Execute code after an ALTER that changes a data type |
| CVE-2026-116330 |
8.8 |
Inject psql commands when restoring a tar created with pg_dump |
| CVE-2026-116318 |
8.8 |
Inject SQL via a constraint name in pg_dump |
| CVE-2026-14661 |
8.1 |
Add content to arbitrary files with dblink and subscriptions |
| CVE-2026-116313 |
7.2 |
Execute code as the user running vacuumdb |
| CVE-2026-116303 |
3.1 |
Bypass the EXECUTE privilege on conversion functions |
The most uncomfortable case in the table is CVE-2026-6460. An ALTER that changes a field's type leaves the previous value with the new tag, and from then on the engine reads a memory address that does not match what it believes it has. Exploitation needs nothing exotic, only permission to create and modify objects.
Client tools are included too
Three of the most serious flaws are not in the server. pg_dump allowed SQL injection through a constraint name and its tar mode generated a restore.sql capable of executing psql commands, two routes for a backup to become the entry vector. Anyone who restores dumps from databases where other users can create objects should also update the client binaries.
The rest of the package touches utilities that are launched from cron without much thought. reindexdb did not properly quote index names, vacuumdb with the incomplete statistics option allowed a database owner to execute code as the user running the tool, and dblink with logical subscriptions could add content to arbitrary files on the system.
That detail changes the priority for a small team. The database tends to be patched carefully and the tools are forgotten, even though they are the ones with permissions over the file system and the ones run with credentials stored in scripts.
Reindex and redo the full backup
Installing the binaries does not close out the work. The batch fixes bugs that could leave indexes corrupt, so it is worth rebuilding suspicious GIN indexes, GiST indexes on pg_trgm with gist_trgm_ops and B-tree indexes on bpchar columns declared without a length. A table returning lost block errors needs a full REINDEX TABLE before the incident can be considered resolved.
Backups also require attention. Anyone using incremental backups should take a new full backup and base the following ones on it, because several of the fixed flaws could silently produce corrupt backups. A backup that fails to restore is discovered at the worst possible moment.
None of this appears in the headline of the announcement and all of it appears in the changelog, a long document worth at least skimming when the installation supports third-party services. Updating your own server has something of household maintenance about it: the a review of what to deploy and what to upgrade in 2026 serves as a reminder that the patch is only part of the job.
The 22-year-old flaw that surfaced with an AI agent
CVE-2026-6460 came with a story of its own. Researcher Mehmet Ince published on November 9, 2026 the third part of his series on managed PostgreSQL risks, in which he recounts how he reached the type-confusion primitive that had been sitting for some 22 years inside the engine. The starting point was not the core, but a badly converted extension that forced him to study how the engine decides what each value is.
In his account, an AI agent served as his teacher to walk through the engine's internals one question at a time, and the idea that connected everything came from a conversation seemingly unrelated to the problem. The detail matters less than the conclusion: the class of flaw was not new, it had simply not been found.
Branch 14 is left without patches
14.25 closes a branch that saw the light of day on September 30, 2021 and whose expiry date was November 12, 2026, five years later, as that same policy sets out. Staying on it is not just an aesthetic decision: from December, any new flaw affecting that version will be left without an official fix, and the project does not maintain bespoke patches for closed branches.
Migrating to 15 or 16 requires pg_upgrade and a real maintenance window, with testing of extensions and the heaviest queries. That work competes poorly against the urgency of the day, and is the reason many installations reach end of support with years of accumulated delay.
A patch that does not fit the monthly window
The usual cycle for a small team is monthly and that of this batch is weekly. Thirty-two 8.8 flaws that allow code execution with system privileges cannot be left waiting for the next change meeting, and even less so when exploitation requires nothing more than a valid connection and permission to create objects.
The other lesson is about inventory. A patch that forces server and client to be updated at the same time, and that also requires reindexing and redoing backups, punishes those who do not know for certain which version is running on each machine. The .NET patch calendar for this same November points in the same direction.
The twelve CISA entries in ten days confirm the pattern, and leave a practical consequence for those who administer databases: the pace of fixes is no longer set by the vendor, but by whoever finds the flaw first.