PostgreSQL 19 is out; its commit log credits AI tools or AI labs in 28 of 90 CVEs fixed since feature freeze
PostgreSQL 19 came out on November 19. At the end of October Robert Haas warned that it would miss the usual release window because the committers were busy fixing bugs that people had found with language models (our report). PostgreSQL 18 shipped on September 25, 2025, so this release arrived eight weeks later in the year.
The release announcement is about REPACK (CONCURRENTLY), the pg_plan_advice extension and WAIT FOR LSN, and doesn't mention the delay. The release notes don't mention the three features pulled in August and September either (MERGE/SPLIT PARTITION, SQL/PGQ and UPDATE/DELETE FOR PORTION OF). They are simply not in 19.
To see where the extra weeks went, I cloned the postgres repository, repeated Haas's count, and carried it forward to the release.
The last two months
Haas's numbers hold up. For April 8 to September 25 I get 1,040 commits to master in 2025 and 1,467 in 2026, 731 and 1,176 on the release line, and 246 and 433 on the previous stable branch. His figures (1,037/1,443, 724/1,169, 248/433) are within 2% of mine; the difference is probably where the day boundaries fall.
The bigger change came after his window closed:
| PostgreSQL 18 | PostgreSQL 19 | |
|---|---|---|
| Betas | 3 | 4 |
| Release candidates | 1 | 2 |
| RC1 tagged | September 1, 2025 | October 12, 2026 |
| Commits from RC1 to the .0 tag | 62 | 340 |
…with a Security: CVE trailer |
0 | 84 |
| Release | September 25, 2025 | November 19, 2026 |
PostgreSQL 18 went from RC1 to release in three weeks. PostgreSQL 19 needed a second release candidate, and 293 commits landed between the two RCs. Eighty-six were committed on one day, November 9, the Monday before the quarterly security release; all 84 security commits came before RC2. Back-branch work grew as well: from September 25 to November 19, REL_18_STABLE took 324 commits, against 103 on REL_17_STABLE over the same weeks of 2025.
Ninety-six CVEs in 2026
The November 12 update (announcement) fixed 52 security vulnerabilities and "over 140 bugs" in 18.7, 17.12, 16.16, 15.20 and 14.25, the last release of 14, and shipped 19 RC2. The quarterly updates before it fixed 5 in February, 11 in May and 28 in August. The project's security page lists 96 CVEs for 2026 and 7 for 2025.
In November's list, several CVEs rated 8.8 end with the words "executes arbitrary code": type confusion after ALTER ... SET DATA TYPE, a parallel RECORD typmod use-after-free, a plperl prepared-query use-after-free.
Who reported them
PostgreSQL commits name their reporters in Reported-by: lines. I took every commit since April 8, 2026 on master, REL_19_STABLE and REL_18_STABLE that carries a Security: CVE-2026-… trailer and grouped the reporters by CVE. That covers 90 CVEs (the February five were fixed before the window, and one August CVE has no such trailer on these branches). Twenty-eight of the 90 name an AI system, or an AI lab's security team or program, among the reporters:
- "Anthropic OSS program": 16 CVEs;
- "Claude and Ada Logics": 4, all committed on November 9, among them CVE-2026-120150, a use-after-free in expanded-record domain checks rated 8.8 (commit);
- "Ben Morris in collaboration with Claude and Anthropic Research": 3, and "Calif.io in collaboration with Claude and Anthropic": 1;
- OpenAI Codex Security (twice as "Amy Burnett (OpenAI Codex Security)"): 3, and "OpenAI Security Research Team": 1.
The share grew through the year: 1 of 11 CVEs in the May batch, 5 of 27 in August, 22 of 52 in November.
Many bugs were reported more than once. Thirty of the 90 CVEs list two or more reporters. CVE-2026-6460, the ALTER ... SET DATA TYPE type confusion, lists 18: OpenAI's Security Research Team, Nicholas Carlini (at an anthropic.com address), committers Heikki Linnakangas, Andres Freund and Peter Geoghegan, and people from Microsoft, PlanetScale and Ada Logics among them. The to_char heap overflow, CVE-2026-14669, lists 11, including OpenAI Codex Security and Wiz's Hillai Ben Sasson (commit). In the same window of 2025, six CVEs were fixed with such a trailer, and only one had more than one reporter.
Some fixes went well beyond a patched call site. For CVE-2026-6460, Andres Freund committed "Introduce attribute versioning to prevent unsafe deforming" (commit), back-patched to version 14. It marks composite values and versions named composite types, so that reading a row with a descriptor of the wrong shape raises an error instead of crashing or being exploited. The commit message calls it "a significant piece of infrastructure grafted onto the existing code" and lists the costs: extensions need a rebuild for full protection, and downgrading a minor version after DDL can cause errors. A follow-up the same day is titled "Work around broken TimescaleDB code".
A Reported-by: line names a person or a program, not the tools used. Sixty-two of the 90 CVEs credit only people by name, and some of them may have used models as well; the log can't tell. So 28 is what the project explicitly credited to AI, not an estimate of the AI share. Haas's explanation fits the log, and the explicit credit is concentrated in the November batch that landed between the two release candidates, but the log doesn't prove it.
PostgreSQL is enrolled in Anthropic's scanner
On October 19 Noah Misch committed a Dockerfile at src/tools/oss-scanner/ for anthropics/oss-scanner (commit), noting "Claude wrote some of this." On November 6 the scanner repository added projects/postgresql/project.yaml: it scans master and sends reports to security@postgresql.org. It also names src/tools/oss-scanner/threat_model.md, which I couldn't find in the postgres repository (master, REL_19_STABLE, REL_18_STABLE and the full history) or in the scanner's projects/postgresql/ directory, the two places the scanner's README allows for one. The scanner's README says its reports are model-generated, not reviewed by a human, and come with a reproducer and a proposed patch where available. We looked at how that went for curl in an earlier analysis. The commits credit "Anthropic OSS program" without saying which pipeline produced each report, so I wouldn't tie those 16 CVEs to the scanner itself.
Haas asked in October how the project would ever catch up with forty years of bugs that models can now find. The 2026 numbers don't answer that, but they show how fast the reports kept coming: 11 CVEs in May, 28 in August, 52 in November. And since November 6 PostgreSQL has an entry in Anthropic's scanner that points at master.
- PostgreSQL 19 Released! (Nov 19, 2026)
- PostgreSQL 19.0 release notes
- PostgreSQL 18.7, 17.12, 16.16, 15.20, 14.25 and 19 RC 2 released
- PostgreSQL 18.6, 17.11, 16.15, 15.19, 14.24 and 19 Beta 3 released
- PostgreSQL security information
- postgres: Avoid using a dangling pointer in expanded-record domain checks (CVE-2026-120150)
- postgres: Guard against overlength time zone abbreviations in to_char() (CVE-2026-14669)
- postgres: Introduce attribute versioning to prevent unsafe deforming (CVE-2026-6460)
- postgres: Add Dockerfile for github.com/anthropics/oss-scanner
- anthropics/oss-scanner: projects/postgresql/project.yaml
Comments
No comments yet. Be the first.