Red Hat disclosed two critical flaws in Quay, its container registry, on November 11, 2026, and shipped fixed builds for eight Quay 3 branches. Both bugs are about tokens, and both end the same way: someone who shouldn’t be able to can read and overwrite private images. Red Hat’s Dave O’Connor found both. According to the bug report, the logic behind the first one has been in Quay since its first public commit.
CVE-2026-123710: planting your own signing key (CVSS 9.8)
Quay signs the JWTs that authorize pushes and pulls with a set of approved service keys, managed by a key server at /keys. To rotate a key, the caller presents a JWT, and Quay checks that it’s signed by an approved key and has the expected audience and issuer.
What Quay didn’t check is whether the token was meant for key rotation. The anonymous registry token from /v2/auth is signed by an approved key, and the caller picks its audience. So one anonymous request can install an attacker-controlled key, approved and with no expiry date. Quay then accepts tokens signed with it for any repository.
What an attacker gets:
- Read every private image, plus any credentials baked into the layers.
- Push or retag images, which then flow into every cluster and pipeline that pulls them.
- Persistence. The key survives restarts, password resets and robot token changes, and it can be used to delete the legitimate keys so normal pushes and pulls fail.
Anonymous access is on by default in Quay. Red Hat is clear that turning it off doesn’t protect you, because any valid account can do the same thing. Affected setups include standalone installs, the Quay Operator on OpenShift, and the mirror registry for Red Hat OpenShift. That covers internal registries too, if untrusted clients can reach /keys.
CVE-2026-110697: an underscore that leaks robot tokens (CVSS 9.9)
The second bug is a classic. The robot-account listing code builds a SQL LIKE pattern from the namespace name without escaping wildcards, and in LIKE, _ matches any single character. A user who can create a namespace with an underscore in the right position gets back another tenant’s robot accounts, with decrypted tokens. Those tokens let the attacker read, push or overwrite the victim’s images. Where Quay is OpenShift’s integrated registry, that’s a direct way into CI/CD.
Deleting the attacker’s namespace also deletes the matching robot accounts, which can break whatever depended on them, so plan the cleanup carefully.
Fixed builds
| Branch | CVE-2026-123710 | CVE-2026-110697 |
|---|---|---|
| 3.9 | 3.9.28 | 3.9.28 |
| 3.10 | 3.10.28 | 3.10.28 |
| 3.12 | 3.12.24 | 3.12.24 |
| 3.14 | 3.14.10 | 3.14.10 |
| 3.15 | 3.15.9 | 3.15.9 |
| 3.16 | 3.16.7 | 3.16.7 |
| 3.17 | 3.17.7 | 3.17.7 |
| 3.18 | not confirmed | 3.18.3 |
Quay.io is already fixed. The upstream patch adds the missing rotation check, requires an expiry on keys added through the key server, stops callers from choosing a token’s audience, and escapes the robot search.
What to do, in order
- Preserve evidence first. Export Quay audit logs before you touch keys.
- Block
/keysat the proxy or load balancer for everything except trusted internal components. - Hunt for planted keys. Look for
service_key_rotateandservice_key_deleteaudit events whereexpiration_dateis null, and remove any approved key you don’t recognize. Until the fixed build is running, this cleanup won’t stick, because the same anonymous request can add another key. - Upgrade to the fixed release for the branch you actually run.
- For CVE-2026-110697, rotate robot tokens for any namespace an underscore name could match, limit who can create namespaces containing
_, and watch for robot pulls and pushes that don’t fit normal use. - Assume images may have been tampered with if you find a rogue key. Rebuild from source or check against known digests and signatures before you trust what’s in the registry.