What the advisory says, and what it does not
CVE-2026-117691, tracked by the GitHub advisory GHSA-c85j-3hq2-q6wj, affects Keycloak's SAML HTTP-Redirect binding. The server decodes percent-encoded parameter names, for example %53AMLResponse, while signature verification only recognizes literal names. The attacker first sends an unsigned message under the encoded name, then a legitimately signed one. The signature check looks at the second, the server processes the first.
Red Hat draws the most serious consequence: takeover of any federated account. The advisory is dated November 5, 2026. The CVSS 3.1 score is 9.1, with a network vector and no privileges or user interaction required, and the weakness is classified CWE-347, improper verification of a cryptographic signature.
Red Hat adds a prerequisite: the attacker needs an account at the external identity provider, to obtain a genuinely signed redirect message, for example a logout response. The GitHub advisory itself lists neither affected versions nor a workaround. The rest of this guide therefore separates what is established from what is deduced.
Who is exposed
- Keycloak as a SAML identity broker — the project ticket (keycloak#54714) speaks of a "broker login as any user". The case described is a realm that federates an external SAML identity provider: an enterprise client, a partner, a third-party directory. This is the scenario to handle first.
- An account at the external provider — the attack reuses a signed message, so it needs an account, even an unprivileged one, at that provider. The more open its registration, the wider the exposure.
- OIDC only — the sources read describe only the SAML HTTP-Redirect binding and do not mention OIDC. A realm with no SAML provider falls outside the scenario described, but that does not excuse skipping the upgrade: the 26.8.2 notes announce 33 security fixes.
- Keycloak as the SAML provider for your applications — no source says whether SAML clients are affected. The 26.8.2 notes list other SAML fixes. Do not conclude there is no risk: update.
- Red Hat build of Keycloak — the NVD data marks the 26.6 branch affected before 26.6.9. Red Hat Single Sign-On 7 is listed by Red Hat as not affected.
- Observed exploitation — as of November 18, 2026, the CVE is not in the CISA catalog of exploited vulnerabilities, and its EPSS score is below 1%. This is not a grace period: the severity score remains 9.1.
Which version to install: what the sources say to each other
For the community project, the fix is in 26.8.2, released on November 5, 2026. In the keycloak#54714 thread, users ask which older versions are affected. The only answer is a pointer to the release/26.6.9 and release/26.4.18 labels in the repository. Why not 26.7: the question was asked on November 9, 2026, with no maintainer reply in the thread.
The sources do not fully agree. Red Hat released Red Hat build of Keycloak 26.6.9 on November 7, 2026, with advisory RHSA-2026:96040. But for the community project, the GitHub release list and the quay.io image tags show neither 26.6.9 nor 26.4.18 on November 18, 2026: the latest published 26.6 is 26.6.4, the latest 26.7 is 26.7.5.
Practical consequence. With the community image, target 26.8.2. With Red Hat build of Keycloak, 26.6.9 or, according to the project tracker, 26.4.18: confirm with your vendor. On 26.7, no fix is announced: treat it as affected. Before 26.4, the advisory establishes nothing, and we conclude nothing: no fixed version is announced there, which is one more reason to migrate.
Keycloak branches against CVE-2026-117691
Scroll the table
| Branch | Status against the CVE | Fixed version |
|---|---|---|
| 26.8 (community) | Not fixed before 26.8.2 | 26.8.2, released November 5, 2026 |
| 26.7 (community) | Latest published: 26.7.5, no fix or backport announced | Move to 26.8.2 |
| 26.6 (community) | Latest published: 26.6.4, no fix | Move to 26.8.2 |
| 26.6 (Red Hat build of Keycloak) | Affected before 26.6.9 according to NVD | 26.6.9, advisory RHSA-2026:96040 |
| 26.4 | Backport announced by the project tracker, no advisory read | 26.4.18 per the tracker: to be confirmed |
| Before 26.4 | Not established by the advisory | None announced: migrate to 26.8.2 |
| Red Hat Single Sign-On 7 | Not affected, according to Red Hat | Not applicable |
Reading the version actually deployed
Record what is running before changing anything, using two routes that cross-check each other.
Admin console: sign in to the master realm, open the dashboard and then the Server Info tab, which describes the server and its enabled features. Command line: the -V or --version option of kc.sh prints the version. Under Docker, the command is docker compose exec keycloak /opt/keycloak/bin/kc.sh --version, adapting the service name.
The tag in the Compose file tells you what you asked for, not necessarily what is running: a floating tag such as 26.8 changes on every docker compose pull. And if your image is built with kc.sh build, the version of the built image is what counts. Finally, in each realm, list the SAML identity providers (Identity providers menu): that tells you whether you are in the scenario described.
Updating Keycloak under Docker Compose
Note the current version and state
Run
docker compose exec keycloak /opt/keycloak/bin/kc.sh --version, then note the realms, SAML identity providers and clients that serve your customers. This list becomes the test grid at the end.Back up the database and the configuration
Back up the database with your engine's tool, for example
docker compose exec -T postgres pg_dump -U keycloak keycloak > keycloak-avant-maj.sql, adapting service, user and database. Also copy the Compose file,conf/, the themes andproviders/. The upgrade guide asks you to back up the database and the old installation.Read the migration changes
The upgrade guide lists changes version by version. Read all those between your version and the target, not just the latest: the 26.8 section changes access control on the admin API and the handling of SAML assertions.
Change the image tag
In the Compose file, switch to
image: quay.io/keycloak/keycloak:26.8.2: an exact tag, neverlatest. Custom image: change theFROMline of the Dockerfile, then rundocker compose build --pull. Official image:docker compose pull keycloak.Stop the old version
Run
docker compose stop keycloak. Rolling updates are documented only for patch releases, since 26.6.0; for a minor version change, stop all nodes before migrating the database.Start and let the database migrate
Run
docker compose up -d keycloakthendocker compose logs -f keycloak. The database migrates by itself on first start. The default timeout is 30 minutes; for a large database, add--transaction-setup-timeout=60mto the start command. Beyond 300000 records, some indexes are created in the background after startup.Retest themes, providers and logins
Compare your custom theme templates with those of the new version, then replay the grid noted in step 1: one login per SAML provider, a logout, an admin API client, an existing federated account.
Plug in watching and monitoring
Add Keycloak's login URL to an uptime probe, for example the Uptime Kuma template from the catalog on the same VPS, and follow the keycloak.org blog for the next patch release. An SSO is updated at the pace of its advisories, not at the pace of your projects.
A backup that restores
After the migration, the database schema is no longer compatible with the old server, and Keycloak cannot undo the migration. To roll back, you must restore the old installation, then the database from the backup. Restore it once on a test database before upgrading production: a backup never restored is only a hypothesis.
If the update has to wait a few days
Let us be clear: there is no official workaround. Red Hat states that no measure meets its criteria for ease, deployment and stability, and the GitHub advisory lists none. The leads below reduce exposure, they fix nothing. They are for holding on a few days, not for postponing the upgrade.
Reducing exposure in the meantime
- Restrict registration at the external provider — the attack requires an account there. Auditing its accounts and limiting new sign-ups reduces the surface. This lead appears in a Red Hat note dated November 6, 2026 that is marked as AI-generated, and it states that it does not fix the flaw.
- Disable the SAML provider concerned — in Identity providers, the enable switch cuts logins through that provider. This is our deduction, not an instruction from the advisories: test it on a staging realm, and measure the impact for that client's users.
- Do not rely on "HTTP-POST Binding Response" — according to the documentation, this setting chooses the binding Keycloak uses to answer requests from the external provider. Nothing indicates that it stops Keycloak from accepting a redirect message.
- Do not promise an nginx filter — you can look for requests with an encoded parameter name, but no source validates a complete blocking rule, and encoding can touch any character of the name. Use the reverse proxy to detect, not to feel protected.
- Fix the upgrade date — a dated maintenance window, the backup made and tested, a named owner. The risk of an unpatched SSO is measured in days.
- Inform the clients concerned — if you run their SSO, say which flaw, which version, which upgrade date. Without adding exploitation details.
Checking the fix, then looking for traces
Verify: read the version again in Server Info and with kc.sh --version; it should show 26.8.2 for the community image. Then replay your functional tests: login through each SAML provider, logout. Do not try to replay the attack on production: a forged message on a live system is an impersonation attempt. If proof is required, build an isolated staging realm with test accounts.
Looking for traces: the attack goes through a GET request with an encoded parameter name. In the reverse proxy access logs, a search such as zgrep -iE '[?&][a-z]*%[0-9a-f]{2}[a-z]*=' /var/log/nginx/access.log* finds parameter names containing an encoded character. A hit is a clue, not proof, and no hit is not proof either: rotated logs, unlogged parameters, another proxy upstream.
On the Keycloak side, user events are not kept by default; only errors are. Enable Save events in Realm settings, Events, to keep history from now on, and review unusual federated logins: time, IP address, account. Before you enabled it, nothing was recorded: what you do not find may simply never have been written.
Troubleshooting after the version jump
Moving to 26.8 changes several behaviors, listed in the upgrade guide. These are the ones that most often look like an outage. The guide quotes no startup error message: the symptoms below are the ones it describes.
Symptoms and remedies, from the upgrade guide
- 401 errors from the admin API — tokens must carry the new
kc-admin-apiscope. Custom clients and service accounts must be assigned it;admin-cliand the console are updated automatically. During a rolling upgrade, temporary 401 errors are possible. - SAML login that no longer works — assertions must carry a bearer subject confirmation, and ProxyRestriction conditions are rejected by default. On the adapter side,
entityID="idp"must be replaced with the realm URL. - Automatic account linking refused — it now requires a verified email address or the Trust Email option on the provider. Federated users who used to link on their own may suddenly see a linking screen.
- Database or CPU under more load — login failures are stored in the database by default, which increases connections and CPU.
--features=login-failures:v1restores in-memory storage. - Migration taking too long — the default timeout is 30 minutes, adjustable with
--transaction-setup-timeout. If the stop happened mid-migration, restore the backed-up database before restarting. - Users logged out — with
persistent-user-sessionsdisabled, all sessions except offline sessions are lost, and login flows in progress must be restarted. Warn users before the maintenance window.
Official documentation
The following links are the sources of this guide; reread them before acting.
GitHub advisory: https://github.com/advisories/GHSA-c85j-3hq2-q6wj
26.8.2 release notes: https://www.keycloak.org/2026/11/keycloak-2682-released
Project tracker: https://github.com/keycloak/keycloak/issues/54714
Upgrade guide: https://www.keycloak.org/docs/latest/upgrading/index.html
Red Hat CVE page: https://access.redhat.com/security/cve/CVE-2026-117691
Red Hat advisory 26.6.9: https://access.redhat.com/errata/RHSA-2026:96040
For an SSO you run yourself, the Keycloak entry in the ServOrbit catalog offers a ready-to-deploy VPS template, with PostgreSQL and HTTPS, to which the procedure above applies as is.
servorbit.com/ma-en/blog/keycloak-cve-2026-117691-saml-signature-bypass-update
