Red Hat disclosed two critical flaws in Red Hat Quay on November 11, 2026. Quay is the container registry behind many OpenShift and Kubernetes estates. CVE-2026-123710 (CVSS 9.8) lets an attacker get a key they control trusted as a registry signing key. The attacker then mints registry tokens that read or overwrite any repository, private ones included. CVE-2026-110697 (CVSS 9.9) breaks tenant isolation in multi-tenant deployments: one tenant can list and steal another tenant’s robot account tokens.

Red Hat says the flawed key-rotation logic has existed since Quay’s first public commit. Fixed builds are available for eight Quay 3 branches, and Quay.io has already been patched. We have found no confirmed in-the-wild exploitation so far. But anyone who can push to your registry controls what your clusters pull and run, so treat this as a supply-chain-grade risk.

CVE-2026-123710: forging your way into the key server

Quay’s key server (/keys) manages the service keys used to sign registry JWTs. Rotating a key through replace_service_key requires a JWT signed by an already-approved key. The check validates only the signature, audience and issuer. It does not check that the token was meant for key rotation.

The registry auth endpoint will issue a signed token to an anonymous client. That token is signed by an approved key, so it passes the rotation check. The audience also comes from the client-supplied service parameter and is not pinned. The attacker submits their own public key as the “replacement.” Quay then:

  1. copies the signer’s approval onto the attacker’s key, and
  2. stores it with no expiration if the expiration parameter is left out.

From then on, Quay accepts tokens the attacker signs with that key, for any repository and any scope. That means pulling private images, pushing backdoored layers over existing tags, and deleting content.

Preconditions: /keys must be reachable by the attacker. Red Hat’s advisory describes the attack as unauthenticated when anonymous access is on, and anonymous access is on by default. Reporting also notes that with anonymous access off, any valid account can still get a usable token. Do not count on disabling anonymous access as a fix.

CVE-2026-110697: LIKE wildcards across tenants

The robot-account listing function (list_entity_robots in data/model/user.py) builds a SQL LIKE pattern from the namespace name without escaping wildcard metacharacters. Robot accounts are named namespace+robot, and _ matches any single character in LIKE. So an attacker who can create an organization or user namespace can pick a name containing _ that also matches a victim namespace.

The result is that the attacker can list, and steal, the victim tenant’s robot tokens. With those tokens they can read, push or overwrite the victim’s images. Red Hat’s Bugzilla entry (CWE-89) also lists robot takeover by renaming a namespace and deletion of a victim’s robots. No victim interaction or elevated privileges are needed.

Affected and fixed versions

Red Hat lists all Quay 3 and Quay 5 releases as affected. The upstream fix is commit 094cfb024f01. It adds a dedicated rotation claim to key-server tokens, pins the registry audience, requires an expiration on new keys, and escapes the robot LIKE pattern. That one commit addresses both CVEs.

Fixed releases reported for the 3.x branches:

BranchFixed in
3.93.9.28
3.103.10.28
3.123.12.24
3.143.14.10
3.153.15.9
3.163.16.7
3.173.17.7
3.183.18.3

Check these numbers against Red Hat’s errata before you rely on them. Early on, the CVE record still showed no fix, and the eight advisories do not name CVE-2026-110697 even though the CVE page lists the same releases as fixed. The affected images are quay/quay-rhel8, quay/quay-rhel9, and upstream Project Quay.

What to do now

  • Upgrade every Quay deployment to the fixed build for your branch. That includes operator-managed instances on OpenShift and mirror registries for disconnected clusters.
  • Block /keys at the ingress, route or load balancer until you are patched. Allow only the Quay components that need it.
  • Audit service keys. In the superuser panel or the database, remove any approved key you don’t recognize, especially one with no expiration. Search audit logs for service_key_rotate and service_key_delete events with a null expiration. Removing a rogue key only helps after the upgrade: before it, an attacker can add another the same way.
  • Rotate robot tokens in multi-tenant registries, and look for recently created namespaces with _ in their names.
  • Check image integrity. Compare digests of critical tags against your CI build records or cosign/Sigstore signatures. Look for pushes from unexpected tokens since the deployment went live. If you find unexplained overwrites, treat it as a supply-chain compromise: rebuild from source and redeploy.
  • Restrict network access to Quay to trusted networks, and turn off anonymous access where you don’t need it, as defense in depth.

References