GPU VulnDB

Database/Control plane, storage & DevOps

Keycloak: authenticated user can exhaust server memory via unbounded Prometheus metric labels

CVE-2026-16100Control plane, storage & DevOpscurated

Impact

When user-event metrics are enabled, Keycloak records raw error text from failed account operations as Prometheus label values, and that text embeds user-supplied input such as a nonexistent client ID. Any user who can authenticate can loop over unique client IDs and mint an unbounded number of distinct metric series until the JVM heap is exhausted and the server crashes or stops answering. On a GPU fleet this is an availability hit on the identity layer rather than on the GPUs: jobs already running keep running, but if Keycloak fronts the scheduler UI, the Kubernetes API OIDC path, the registry or the observability stack, tenants cannot get new tokens and operators cannot log in to respond. Recovery is a restart, so the blast radius is bounded, but a scripted attacker can re-trigger it as fast as the service comes back.

Who can reach it

Any account that can authenticate to the realm and reach the account/client endpoints - network-reachable, low privileges, no admin rights, no user interaction. Not exploitable unauthenticated, and not exploitable at all if user-event metrics are disabled.

What to do

Red Hat shipped errata RHSA-2026:50848 and RHSA-2026:50849 for the Red Hat build of Keycloak; apply the updated build and restart the Keycloak server or roll the Keycloak pods. No node drain or reboot is involved - this is a daemon restart of the IdP, which does log out in-flight authentication attempts, so schedule it alongside other control-plane work. If the update cannot be applied immediately, disabling user-event metrics removes the code path that writes the attacker-controlled labels. The record does not name a fixed upstream version beyond the affected 26.6.x line.

References

Related entries

All Control plane, storage & DevOps entries

This entry is curated: imported from vendor advisories with machine assistance, not yet individually verified. Confirm against your vendor's advisory before acting, and report anything wrong.