, on
dash and on bash. It never sends a request, never runs an exploit and never
It finds every MySQL server program on the machine (running and merely installed,
containers included when their processes are visible) and the classic client
programs, identifies the product (Oracle MySQL, MariaDB or a fork), judges the
version **against Oracle's affected list**, and, for a server inside the list,
**Read the section first: the public sources are thin, and this audit says
What the sources state (the CVE record published by Oracle as the CVE Numbering
Authority, NVD, and the CWE added by CISA):
- **Component:** "Client programs", in MySQL Server and in MySQL Cluster.
- **Description:** an easily exploitable vulnerability that allows an
**unauthenticated attacker with network access via multiple protocols** to
compromise the product. Successful attacks can result in unauthorized create,
delete or modify access to data, and unauthorized access to all accessible data.
- **Score:** CVSS 3.1 base score **9.1**,
`AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:N` (no availability impact).
- **Weakness:** CWE-287, Improper Authentication (added by CISA).
- **Affected versions, as Oracle lists them:**
| MySQL Server | 8.4.0 to 8.4.13, 9.7.0 to 9.7.4, and 26.10 |
| MySQL Cluster | 8.0.0 to 8.0.50, 8.4.0 to 8.4.13, 9.7.0 to 9.7.4, and 26.10 |
- **Not in the list:** MariaDB (a different product). MySQL **8.0 Server** is not
in Oracle's Server list (8.0 *Cluster* is). Every other train (5.7, 8.1 to 8.3,
9.0 to 9.6, Cluster 7.x, ...) is not listed either.
- **Status as of 2026-11-19:** no exploitation reported, not in the CISA KEV list.
What **no source states**, and this script does not invent:
- **Fixed versions.** Neither the CVE record nor any other source used here names
one. The listed ranges end at 8.4.13 and 9.7.4, which only *suggests* that later
releases are not affected; that is a guess, not a statement. NVD has no CPE
- **What the flaw is**, why a bug in "client programs" has a network attack vector
("multiple protocols"), what the attack path is, and whether any workaround
- **The exact version form of the new year-style release "26.10".** The script
reads `26.10` and `26.10.N` as the listed `26.10`; if Oracle ships it in another
form (for example `26.10.0-1`), that is read the same way, and anything that
command line tools, so the affected list above comes from the CVE record. Check
the advisory and Oracle's patch availability document yourself.
# review it first, then run as root (other users' /proc entries need root)
sh cve-2026-114952-mysql-audit.sh --quiet # verdict only
sh cve-2026-114952-mysql-audit.sh --no-banner # text report without the banner
On a remote host, nothing is copied to disk:
`MITIGATED` > `PATCHED`); `NOT_APPLICABLE` only when every instance is. A running
`mysqld` whose binary path cannot be read lowers a clean result to `UNKNOWN`.
1. **Instances.** Running `mysqld`, `mysqld-debug` (and MariaDB's `mariadbd`)
processes through `/proc`; a process in another mount namespace is marked as a
container and is never executed. Installed server binaries in the usual places
(`/usr/sbin`, `/usr/libexec`, `/usr/bin`, `/usr/local/mysql/bin`,
`/opt/mysql*/bin`, ...) and the classic client programs `mysql`, `mysqldump`,
`mysqladmin`, `mysqlcheck`, `mysqlimport`, `mysqlshow` and `mysqlbinlog`,
because the vulnerable component is named "Client programs". Instances are
deduplicated by resolved path and container (two containers with the same
program path stay separate, as do a host and a container).
2. **Product and version.** From `PROGRAM --version` when it is safe to run, else
from the strings inside the file (never executed). `MariaDB` in the banner:
`NOT_APPLICABLE`. `Percona`, `XtraDB`, `Galera`, `TiDB`, `Aurora`: `UNKNOWN`
("a fork; Oracle's list names Oracle MySQL, check the fork vendor"). A banner
with a version and no fork name (`MySQL Community Server`, `MySQL Enterprise`,
a source build, a distribution tag such as `(Ubuntu)`) is read as Oracle code.
3. **Server or Cluster.** Cluster when the banner says so (`-cluster` in the
version, `MySQL Cluster`, the word `NDB`) or when `ndbd`, `ndbmtd` or `ndb_mgmd`
is installed to the program or in a usual program directory. It matters:
a Cluster node at 8.0.50 is in the list, a plain Server at 8.0.50 is not.
4. **Package.** `dpkg`, `rpm` or `apk` ownership of the binary, and the package
changelog for the CVE id (`CVE-2026-114952`).
5. **Exposure of a server inside the list.** The sockets the process really
listens on: its socket descriptors (`/proc/PID/fd`, `socket:[N]`) matched to
the LISTEN entries of `/proc/PID/net/tcp` and `tcp6` (the net namespace of the
process; a host process may fall back to `/proc/net/tcp`). Loopback is
`127.0.0.0/8`, `::1` and `::ffff:127.x.x.x`; everything else, including `::`
and `0.0.0.0`, is reachable. This covers 3306, the X plugin port 33060 and any
other port. Also the configured `bind-address`, `mysqlx-bind-address`, `port`,
`mysqlx-port`, `skip-networking` and `mysqlx` of the `[mysqld]` group (text
parse of `/etc/my.cnf`, `/etc/mysql/my.cnf`, `!include` and `!includedir`,
`--defaults-file` and the command line of the process): shown as information,
and used as the verdict only when the live sockets cannot be read.
Apply the **November 2026 Oracle Critical Patch Update** for your MySQL release
and restart `mysqld`. **The fixed version numbers are not published in the sources
[advisory]( and its patch
availability document for the release you run. Update the **client programs** too:
they are the named component. Distribution builds may carry the fix under an older
version number: read the package changelog or the vendor tracker.
**MariaDB is a different product and is not affected** (it is not in Oracle's
**If you cannot update yet:** Oracle names no workaround and no attack path.
Keeping `mysqld` off non-loopback addresses (`bind-address=127.0.0.1` **and**
`mysqlx-bind-address=127.0.0.1`, or `skip-networking`, which also closes the X
plugin port) or behind a firewall is a precaution. It is not a stated mitigation:
because Oracle does not say how the bug is reached, "not reachable from the
network right now" is only what this audit can see.
Per instance, using the running binary when there is one. First the **product and
| MariaDB | not affected (`NOT_APPLICABLE`) |
| a fork (Percona, Galera, ...) or an unidentified product | **unknown** |
| version cannot be parsed or has no patch level (`8.4`) | **unknown** |
| a train Oracle does not list for that product kind (Server 8.0.x, 5.7, 8.1 to 8.3, 9.0 to 9.6, 9.8 and later, Cluster 7.x, ...) | not affected (`NOT_APPLICABLE`): "not in Oracle's affected list". No end-of-life claim is made |
| same train as a listed range but **above its last listed release** (8.4.14 and later, 9.7.5 and later, Cluster 8.0.51 and later) | `PATCHED` with the basis "not in Oracle's affected list; Oracle names no fixed version, so this relies on the list only" |
| inside a listed range (or `26.10`, `26.10.N`) | **affected**, then the package rules below |
**What `PATCHED` therefore means:** the version is above the end of Oracle's
affected list in its train. It is **not** a statement that Oracle fixed it in that
release, because no source says so. Treat it as "not in Oracle's list", and
For an affected version the package decides:
| the package changelog names `CVE-2026-114952` (id must end at a non-digit) | `PATCHED` (backport) |
| Oracle repository package (`mysql-community-*`, `mysql-commercial-*`, `mysql-cluster-community-*`, `mysql-cluster-commercial-*`) | unpatched: judged by version |
| distribution package (`mysql-server`, `mysql-server-core-*`, ...) with no changelog mention | **unknown**: a backport cannot be excluded |
| container image | **unknown**: the image provenance is unknown, pull a newer image |
| not owned by any package manager (tarball, `/usr/local/mysql`) | unpatched: judged by version |
An unpatched server is then judged by exposure: running with at least one
non-loopback TCP listener is `VULNERABLE`; running with only loopback or
unix-socket listeners is `MITIGATED`; not running is `MITIGATED`; if the process
cannot be inspected the configuration decides (a non-loopback bind: `VULNERABLE`,
loopback or `skip-networking`: `UNKNOWN`). A **client program** inside the list
(and not covered by a changelog) is `UNKNOWN` with the reason that the component
is named "Client programs" and Oracle publishes no attack path, so the affected
client binaries are listed and should be updated. A client alone is never
`--json` prints `script_version`, `cve`, `matrix_date`, `verdict`, `exit_code`,
`os`, `instances`, `instance_reports` and `reasons`. Each entry of
| `exe`, `kind` | program path; `server`, `cluster` or `client` |
| `product`, `version` | `Oracle MySQL`, `MariaDB`, `fork: NAME` or `unknown`; the version token read |
| `running`, `pid`, `container`, `binary_replaced` | process facts |
| `package`, `package_vendor` | owning package; `true` for Oracle's repository, `false` for a distribution, `null` when none |
| `list_state`, `list_basis` | `affected`, `newer`, `not_listed`, `unknown` or `n/a`; the basis text |
| `listeners`, `external_listeners` | the TCP listeners found; how many are non-loopback (`null` when they could not be read) |
| `config_bind` | the configured bind and networking options |
| `verdict`, `reasons` | the instance verdict and its reasons |
- **The attack path is unpublished.** "Reachable from the network" is judged from
listening TCP sockets because the CVSS vector is `AV:N`. If the flaw is reached
through a client program that connects *out* to a hostile server, a loopback or
stopped `mysqld` says nothing, which is why affected client programs keep the
result at `UNKNOWN`. A `MITIGATED` verdict is "not reachable right now", not a
- **No fixed versions.** `PATCHED` for an 8.4.14 or 9.7.5 rests on the end of the
affected list. If Oracle's advisory turns out to fix a different set, the rule
(`last_affected` in the script) is the one place to change.
- **"26.10".** The exact version form of the year-style release is unverified. It
is read as `26.10` or `26.10.N`, any patch level, in both product kinds. A later
`26.x` train is "not in Oracle's affected list".
- **Distribution backports.** A distribution package whose upstream version is in
the list reads `UNKNOWN` unless its changelog names the CVE. Ubuntu keeps only
the last ten changelog entries in a binary package, and `rpm -q --changelog`
depends on the package; a package may also fix the bug without naming the id.
- **Containers** are never executed. The version comes from the strings in the
binary (the source tree name, `mysql-8.4.13/`) and the option files from the
container's root. The package database inside a container is not read, so an
affected image is `UNKNOWN`. The listeners come from the net namespace of the
container process; with docker the audit needs `CAP_SYS_PTRACE` (or to run on
the host as root) to read `/proc/PID/exe` and `/proc/PID/fd` of a server that
- **Forks.** Percona, Galera and similar builds are `UNKNOWN`: Oracle's list names
Oracle MySQL only; check the fork vendor.
- **Non-root runs.** A server owned by another user cannot be inspected; the
- **Cluster detection** needs the banner or an `ndb*` program in a usual directory.
A Cluster node whose binary says neither is judged as a plain Server (8.0.x
- **Address decoding assumes a little-endian host** (x86, ARM, RISC-V, ppc64le):
the kernel prints `127.0.0.1` as `0100007F`. On a big-endian machine loopback
listeners would read as non-loopback (a false `VULNERABLE`, never a false clean
- **Strings fallback.** When a binary is not safe to run, the product is guessed
from `MariaDB`, `Percona`, ... text and the version from the source tree name. A
- **A binary is run only when it is trusted:** owned by root (or by the user who
runs the audit), not writable by others or by a group other than root, and the
same for its directory (and for the target and directory of a symlink, as for
`mysql -> mariadb`). Otherwise the version is read from the file.
- Local firewall rules, port mappings of containers and proxies in front of the
sh tests/run_tests.sh # 638 regression tests on a fake root; run as a normal user
sh tests/docker_tests.sh # the same suite under ash, dash, bash 4.2 and 4.3 in
# containers, plus end-to-end tests on real servers
The unit tests run the audit against a fake root with fake `mysqld` and client
programs, a fake `/proc` (fd links, `net/tcp`, `net/tcp6`, `ns/mnt`, `comm`, `exe`,
`cmdline`) and fake `dpkg-query`, `rpm` and `apk`; the test hooks
(`MYSQL_AUDIT_TEST_ROOT`, `MYSQL_AUDIT_TEST_PATH`) are ignored when the script
runs as root. They cover every boundary of every listed range for Server and
Cluster, product identification, loopback decoding for IPv4 and IPv6, the listener
and configuration rules, the package and changelog rule (with a 400 KB padded
changelog), containers, the trust rules for binaries (with a positive control for
every "never executed" check), aggregation and the output formats.
The end-to-end tests need docker or podman:
- real **Alpine** with a running **MariaDB** server and the clients (`apk add
mariadb mariadb-client`): they must come out `NOT_APPLICABLE`, the real `/proc`
and socket mapping must find the listener (0.0.0.0, 127.0.0.1, and
`skip-networking` with none), plus `--json`, `--quiet` and the banner. A second
run without `CAP_SYS_PTRACE` checks that the unreadable server process gives
- the official **Oracle `mysql:8.4`** image with a real running Oracle MySQL,
whatever version it has: the audit must classify it as Oracle's list says (the
expectation is written down in the test, independent of the script), including
the X plugin port 33060 and `skip-networking`. `NO_ORACLE=1` skips it.
The exact version form of 26.10 and the behaviour on a real Cluster install were
**not** tested against a real server (none was available).
- Oracle Critical Patch Update, November 2026 (answers 403 to command line tools):
The full story
This article is one source in a clustered incident — the cluster page carries the summary, timeline and every other outlet covering it.
