CVE-2026-132165: remote code execution through the Drupal installer
CVE-2026-132165 (SA-CORE-2026-017): the Drupal installer can reinstall an existing site. Fixed in 10.6.22, 11.3.22 and 11.4.13. How to close install.php.

Context
On 18 November 2026, the Drupal Security Team published advisory SA-CORE-2026-017, tracked as CVE-2026-132165. Drupal core ships with an installation script, core/install.php, which stays in place once the site is in production. The advisory states that, under specific conditions, an attacker without an account can use it to attempt to reinstall Drupal, which can lead to site takeover or to code execution on the server.
The advisory is rated critical, with a score of 18 out of 25 on Drupal's scale. It was preceded, two days earlier, by a public service announcement that set the release window and asked site owners to reserve time. Drupal keeps that procedure for flaws it expects to see exploited quickly.
What makes this advisory hard to read is what it leaves out. The "specific conditions" are described nowhere. The risk score speaks of uncommon configurations, while the announcement says that no one should assume their site is safe. A site owner cannot decide, from the text, whether they are concerned. This article starts from that observation and explains why we would not try to decide.
We did not reproduce the reinstallation of a site and did not try to bring together the conditions of the flaw. We read the advisory, the announcement, the documentation Drupal updated the same day and the fix in the core repository. We also built six Drupal instances, affected and fixed, to observe what the installer answers from outside and to try the closing measures.
Technical summary
| Item | Value |
|---|---|
| Identifiers | CVE-2026-132165, advisory SA-CORE-2026-017 |
| Component | Drupal core installer, core/install.php |
| Severity | Critical, 18/25 on Drupal's scale; no CVSS score published as of 19 November |
| Risk vector | AC:Complex/A:None/CI:All/II:All/E:Proof/TD:Uncommon |
| Affected versions | Drupal 10 before 10.6.22; 11.0.0 to 11.3.21; 11.4.0 to 11.4.12; end-of-life branches |
| Fixes | Drupal 10.6.22, 11.3.22 and 11.4.13 (18 November 2026) |
| Condition | "Specific conditions", not detailed by the advisory |
| Discovery | Anthropic's OSS Scanner (reference ANT-2026-GMRPK0BU) |
| Exploitation | Proof of concept reported by the advisory; no exploitation reported; not in KEV |
Drupal does not use CVSS but its own scale, derived from NIST. The vector reads as follows: complex access, no authentication, confidentiality and integrity fully compromised, proof of concept available, uncommon configurations. The official label of that last criterion speaks of "module configurations", which makes no sense for an installer. It is the name of the box in the scale, and it should not be read as meaning that the flaw depends on a module.
On 19 November, the CVE identifier still existed only in the advisory. The cve.org database answered that the record did not exist, the NVD returned no result, and no EPSS score had been computed. Tools that feed on those databases therefore do not know this flaw yet, and a silent dependency scanner proves nothing this week.
Our reading of the severity is this. The stated complexity lowers the probability on a given site, it does not lower the stakes: the outcome described is the complete loss of the site. And the measure that removes the risk costs one web server rule. We would handle the subject within the day, without waiting to know on which side the configuration falls. Our usual triage method is described in the article on vulnerability prioritization.
Affected systems
Supported branches
Three branches receive the fix: 10.6, 11.3 and 11.4. Every version of these branches before the fixed version is affected, and the advisory takes the Drupal 11 range back to 11.0.0.
The announcement of 16 November also named two pre-releases, 12.0.0-beta3 and 11.5.0-beta1. The advisory does not mention them. We did not check whether they received the same fix, and a staging site on one of them deserves the same closing measure as the others.
End-of-life branches
Drupal 8 and Drupal 9, branches 10.5 and earlier, 11.2 and earlier are no longer covered. The advisory does not explicitly say they are vulnerable. It says they receive no fix and points their owners to closing the installer.
We would treat them as affected. The lower bound given for Drupal 11 is the very first version of the branch, which points to an old defect and not to a recent regression. That fleet is not empty. Drupal 9 left support in November 2023, Drupal 8 two years earlier, and sites on those branches are still online.
Drupal 7
The advisory does not mention Drupal 7, whose installer sits at the root (/install.php) and whose community support ended in January 2025. One clue is worth noting, though. Pantheon, which maintains Drupal 7 for its customers with Tag1 Consulting, published a version 7.109 the same day, from which the installer has simply been removed.
A hosting provider does not remove a file from a distribution under extended support without a reason. We did not examine the Drupal 7 code and cannot state that it is affected. We would close its installer anyway, since the measure is the same and costs nothing.
Conditions the advisory does not give
The two official texts do not pull in the same direction. The risk score rates the target as "uncommon". The announcement, for its part, contains this sentence: "More than one configuration is affected, so do not assume your site is safe."
Both statements are compatible. Several configurations are concerned, and each of them is rare. But the site owner who wants to know whether theirs is among them has nothing to go on, and that is intended: describing the conditions would amount to guiding the attack.
We read the fix, which is public, and it confirms that several distinct paths were closed. We do not detail them here, for the same reason as Drupal. What that reading changes in our recommendation fits in one sentence: some of these conditions are not settings one chooses once and for all, and a site that is not exposed today may be on another day. Answering "we are not in that case" does not seem to us a position one can hold.
How it works
This section relies on the advisory, on the Drupal documentation and on the fix. It stays at the level of detail of the advisory.
An installer that stays in place
Installing Drupal through the browser goes through core/install.php. The script runs the familiar steps: choice of language and profile, requirements check, database settings, writing of settings.php, creation of the administrator account.
Once the site is installed, the file is not removed. It is part of core, and Composer puts it back at every update. Drupal relies on a check performed at the start of each request: if a site already exists, the installer stops on a "Drupal already installed" page.
The whole security of the installer therefore rests on that check. One can still read, in the drupal.org forums, that a reachable install.php carries no known danger since it refuses to work on an existing site. CVE-2026-132165 ends that certainty: there are situations in which the installer does not recognise the site in place and agrees to start again.
From reinstallation to takeover
Whoever completes an installation chooses the database, and the name and password of the administrator account. If a third party can restart the installation of an existing site, they obtain a Drupal of which they are the administrator, on the victim's server and under its domain name.
From there to code execution, the path depends on what the administrator of a Drupal can do on the targeted installation. The advisory names both outcomes, site takeover and remote code execution, without specifying which is obtained in which case.
From the installer to takeover
- 1
Anonymous request
the attacker calls core/install.php, reachable on most sites.
- 2
Conditions met
in some situations, the installer does not recognise the existing site.
- 3
Reinstallation
the attacker runs the installation as the legitimate owner would.
- 4
Administrator account
they set the credentials of the new site themselves.
Takeover
the site is theirs, with code execution possible depending on the installation.
The role of settings.php
The advisory gives a single technical mitigation: some scenarios are blocked if the web server cannot write to settings.php nor change its permissions. The second half of the sentence matters as much as the first.
Drupal sets settings.php to read-only at the end of the installation. Many site owners stop there. Yet if the file belongs to the account PHP runs under, that account can give itself write permission back whenever it wants. The fix shows that the installer did so by itself in some circumstances, and it no longer does.
The real protection therefore lies in who owns the file. A settings.php owned by a deployment account distinct from the web server's stays read-only whatever happens. On shared hosting, where the same account owns the files and runs PHP, that separation does not exist and the mitigation does not apply.
Fix
The fix carries the same title on every branch, "fix: SA-CORE-2026-017". On 11.4, it is commit e301840 of the drupal/core repository. It changes a single file, core/includes/install.core.inc, over some sixty lines, and the change is the same size on 10.6 and 11.3.
Without going into the conditions, it tightens three points. The check that establishes that a site already exists is made stricter, and the installer now stops when it cannot establish the state of the site, where it used to carry on. An interactive installation is tied to the browser session that started it. Finally, the installer no longer tries to correct the permissions of settings.php by itself.
We find this correction sound, because it does more than plug one hole. The three measures are independent, and each limits what an attacker would do with a defect in the other two. Tying the installation to the session also protects a situation few people think about: an installation started and left unfinished, on a server that is already reachable.
The commit message credits eighteen people, and the advisory names the discovery: Anthropic's OSS Scanner, a tool for automated analysis of open source software. The defect was in a file every Drupal site has exposed for years. Installers, update scripts and maintenance tools shipped with applications are the kind of code that is written once and seldom reread, and we would not be surprised to see other software receive the same examination.
The second advisory of the day
The three versions also fix SA-CORE-2026-018 (CVE-2026-132166), rated moderately critical. It is a gadget chain in core, usable only if another flaw allows untrusted data to be deserialized. The advisory states that it is not directly exploitable.
It does not change the priority, but it settles a practical question: those who would hesitate to apply an isolated fix by hand have a second reason to install the full release.
Exploitation status
No exploitation had been reported as of 19 November. The flaw is not in the CISA KEV catalog in its version of 18 November, which we reread, and we found neither public code nor a detailed technical analysis.
Two elements temper that finding. The risk vector carries E:Proof: the security team knows that a proof of concept exists, even if it is not published. And the announcement warned that exploits might be developed "within hours or days".
The fix is public and short. Anyone who can read PHP can compare the two versions of the file and understand what was closed, as we did. The hard part, for an attacker, will be finding sites that meet the conditions. But probing install.php on a large number of sites costs nothing, and we expect to see this path appear in automated sweeps before the end of the month. That is our own assessment, which nothing confirms yet.
The announcement finally contains an unusual warning: do not forward, and do not act on, an unsolicited email describing a Drupal core vulnerability. A critical alert announced in advance is an opportunity for phishing, and teams that run a Drupal will probably receive fake "emergency fixes" in the coming days.
Checking a site
The installed version
The reliable check is done on the server, at the root of the project:
composer show drupal/core | grep versions
vendor/bin/drush status --field=drupal-versionA version below 10.6.22, 11.3.22 or 11.4.13, depending on the branch, is affected. So is a version of an end-of-life branch, with no fix available.
What the installer answers
The Drupal documentation suggests three requests, to be run from outside:
curl -I https://www.example.com/core/install.php
curl -IL https://www.example.com/install.php
curl -I https://www.example.com/core/install.php/fooAll three must answer 403 or 404. A 200 means the installer is served. On our test instances, a standard Drupal answers 200 to the first and the third, and redirects the second to the first. The third form is not a detail: a blocking rule written on the exact path lets the variant with a suffix through.
Our lab taught us something else. On an installed site, the "Drupal already installed" page publishes the exact core version in the address of its scripts (drupal.js?v=11.4.12). The site itself is far more discreet: its X-Generator header only announces "Drupal 11", and its aggregated scripts carry no version. A reachable installer therefore tells anyone whether the site is up to date, before any attempt. That is one more argument for closing it, including on a fixed version.
This reading is enough to situate a site from outside: installer served, version in range. It does not say whether the conditions of the flaw are met, and no harmless request would. On that basis we wrote a Nuclei template, CVE-2026-132165-flawfence, which is limited to one GET request on the installer and submits nothing. It flags our 11.4.12 and 10.6.21 instances, and stays silent on 11.4.13 and 10.6.22 as well as on a site whose installer is closed.
We have also added a Drupal scan to our free tools. It makes the same reading as the template: the home page, then the installer, without submitting anything to it.
That last case calls for a reservation. A vulnerable site whose installer is blocked shows up in no external check, which is the intended effect. It still has to be updated, because a web server rule is lost more easily than a version.
Permissions on settings.php
ls -l web/sites/default/settings.php
ps -o user= -C php-fpm,apache2,httpd | sort -uThe file must not belong to the account the second command returns, and that account must not be able to write to the sites/default directory. If the owner and the web server account are the same, the 444 mode displayed protects nothing.
Remediation
- Close the installer with a web server rule, today, on every Drupal site, fixed or not.
- Update to 10.6.22, 11.3.22 or 11.4.13.
- Check the owner of
settings.phpand of thesites/defaultdirectory. - Plan the migration of end-of-life sites, for which closing the installer is the only measure available.
Closing the installer
Drupal recommends using the web server configuration in preference to .htaccess. For Apache, in the virtual host configuration:
<Location "/core/install.php">
Require all denied
</Location>For nginx, in the server block:
location ^~ /core/install.php {
return 404;
}On hosting where only .htaccess is available, the equivalent rule works, with a catch: that file is managed by Composer and will be overwritten at the next core update. The documentation describes how to have the lines appended by Composer's scaffold mechanism instead of by hand.
We tried closing the installer through Apache on an 11.4.12 instance. The three verification requests answer 403, the site works normally, and update.php is not affected. That last point is worth checking at home: update.php is used for database updates and must not be swept away by a rule that is too broad.
Removing the file outright is possible, and it is Pantheon's choice for Drupal 7. On Drupal 10 and 11, Composer restores it at every core install or update. It then has to be deleted by a post-install script, which the --no-scripts option bypasses. We prefer the web server rule, which does not depend on how the deployment is launched.
Updating
composer update "drupal/core-*" --with-all-dependencies
vendor/bin/drush updatedb
vendor/bin/drush cache:rebuildSites protected by Drupal Steward, the web application firewall offered by the Drupal Association, are covered against the known vectors. The advisory states that the update is still required, as other vectors may appear. Pantheon, for its part, has blocked the installer on its platform. In both cases the protection comes from a third party and covers what is known today.
End-of-life sites
For a Drupal 8, 9 or an abandoned branch, there is no fix to wait for. Closing the installer and locking settings.php are the only two measures, and they rest on a configuration nothing guarantees over time.
These sites also accumulate every flaw published since the end of their support. This advisory is a good occasion to put their migration back on the agenda, with an argument the previous ones did not offer: the flaw is in a file every site exposes.
Looking for compromise
Drupal has published no indicator. Those that follow derive from the nature of the flaw and from our reading of the fix. We have not checked them against any real attack.
The first signal is in the access logs. On a production site, nobody has a legitimate reason to call the installer, let alone to send it a form:
grep -h 'install\.php' /var/log/apache2/*access* /var/log/nginx/*access* 2>/dev/null \
| grep -v ' 40[34] 'Isolated GET requests will be numerous in the coming days: they are sweeps, ours included if you use our tools. POST requests, or a series of requests carrying step parameters, are of another order and justify a full examination.
A successful reinstallation does not go unnoticed for long, since the content of the site disappears. The traces to look for are a recent modification date on settings.php, database settings that are not the owner's, an unknown administrator account, and a site installation date (the install_time value) later than its actual go-live.
stat -c '%y %U %a' web/sites/default/settings.php
vendor/bin/drush state:get install_timeIf one of these signals is confirmed, the update is no longer enough. Whoever administered the site may have dropped code on it, and the server has to be examined as a machine on which a third party ran commands. The secrets held in settings.php (database credentials, hash salt) have to be renewed.
Timeline
Timeline of CVE-2026-132165
Announcement
16 November 2026
Public service announcement
PSA-2026-11-16 sets the release window and warns of possible exploitation within hours.
Release
18 November 2026, 14
Fixes in the repository
the "fix: SA-CORE-2026-017" commits carry that time on the three branches.
18 November 2026, 17 to 21
Advisory and versions
SA-CORE-2026-017 is published with Drupal 10.6.22, 11.3.22 and 11.4.13.
18 November 2026
Documentation
Drupal's page on securing the installer is updated the same day.
18 November 2026
Hosting providers
Pantheon blocks the installer on its platform and removes install.php from its Drupal 7.
Aftermath
19 November 2026
Status on that date
no record on cve.org or at the NVD, flaw absent from KEV, no public code found.
Frequently asked questions about CVE-2026-132165
What is CVE-2026-132165?
It is a flaw in the Drupal core installer, published as SA-CORE-2026-017. Under specific conditions, an attacker without an account can use core/install.php to attempt to reinstall an existing site, and take control of it.
Which Drupal versions are affected?
Drupal 10 before 10.6.22, Drupal 11.0.0 to 11.3.21 and Drupal 11.4.0 to 11.4.12. End-of-life branches, including Drupal 8 and 9, receive no fix and should be treated as affected.
Is my site in the "specific conditions"?
Nobody can tell from the advisory, which does not describe them. Drupal writes that more than one configuration is affected and that no one should assume their site is safe. The reasonable answer is to close the installer and update without trying to decide.
Is blocking install.php enough?
Blocking removes the known attack path and takes a few minutes. It does not replace the update: a web server rule can disappear during a migration or a change of hosting, and the vulnerable version would be exposed again.
Is the flaw being exploited?
No exploitation had been reported as of 19 November 2026, and the flaw is not in the CISA KEV catalog. The advisory states that a proof of concept exists. Drupal expects exploits to follow quickly, and the fix is public.
Is Drupal 7 affected?
The advisory does not mention it, and we did not check. Pantheon removed the installer from its maintained version of Drupal 7 on the day of the release. We would close /install.php on a Drupal 7 as a precaution.
Should an up-to-date site still close the installer?
Yes. The Drupal Security Team recommends it in every case after installation. The installer page also publishes the exact core version, which the site does not otherwise reveal.
Sources
- Drupal, advisory SA-CORE-2026-017, 18 November 2026
- Drupal, public service announcement PSA-2026-11-16
- Drupal, documentation "Securing the Drupal installer"
- drupal/core repository, commit "fix: SA-CORE-2026-017" (11.4 branch)
- Drupal, advisory SA-CORE-2026-018, 18 November 2026
- Pantheon, release note of 18 November 2026
- CISA, Known Exploited Vulnerabilities catalog
Let’s discuss your external exposure
Request a personalized Flawfence demo and discover your real exposure level.
