// hackerlogs
login+ register
Agentic AILLM AppSecThreat BriefCritical

The July Commit Moved the Key. The Bug Is Still New.

CVE-2026-117574 published at 20:02 UTC on 28 October. Bugzilla 2494861 is still NEW. The 3 July commit changed lookup, and the score note says that was not a fix.

The short answer

CVE-2026-117574 published at 20:02:14 UTC on 28 October 2026. An authenticated OpenShift AI gateway caller can name another tenant's model and receive that tenant's provider API key. Bugzilla 2494861 remains NEW. Commit bac9fb5 on 3 July moved lookup off the URL path. The score note says that change was not a fix. The CVE lists no version.

Key takeaways

  • The primary is CVE-2026-117574, Red Hat, published at 2026-10-28T20:02:14.895Z. CVSS 3.1 is 9.1. Red Hat's word is Important. Bugzilla severity is high. CWE-639.
  • Bugzilla 2494861 opened at 2026-06-30T12:16:08Z. Status is NEW, resolution empty, cf_last_closed null. Last change 2026-10-27T12:46:29Z. From the open to the CVE is 120 days, 7 hours, 46 minutes, 6.895 seconds.
  • Commit bac9fb55 merged at 2026-07-03T21:04:40Z as a feature, PR 359. It is an ancestor of main. The only tag, v0.1.0-rc.1, is 21 April and does not contain it. GitHub releases length is 0.
  • The CVE title says a URL path. The description says the body or a header. The score note says bac9fb5 removed the path lookup and did not add an identity check.
  • The runtime package rhoai/odh-ai-gateway-payload-processing-rhel9 is defaultStatus affected. The e2e package is unaffected. Neither row has a version.

US trends at 22:30 UTC on 28 October were Fofana, Ernie, and Liverpool. Gemini 4 Carbon sat on the same list as a launch. Hacker News led with a fingerprinting note and that launch. X's news tab returned 403. CVE-2026-117574 published at 2026-10-28T20:02:14.895Z.

Critical Critical on the CVSS 3.1 vector, score 9.1, CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:C/C:H/I:L/A:L. Red Hat's classification word is Important. Bugzilla 2494861 says severity high. The flaw is CWE-639. An authenticated caller can draw another tenant's provider API key.

Three panels for the OpenShift AI tenant key: a July lookup change, a CVE with no version, and a bug that is still NEW.

The July commit changed the lookup.#

The primary is Red Hat, product Red Hat OpenShift AI, component ai-gateway-payload-processing. The CVE description says an authenticated remote attacker names another tenant's model identifier in the request body or header. The service retrieves credentials from the shared store before it checks authorization, and it injects the target tenant's API key into the request. The caller can then use that tenant's model quota. The statement says the rating is Important because the caller must already hold a valid user credential for the model-serving gateway, which limits unauthenticated use. This brief does not include a request.

The title and the Bugzilla summary still say tenant and credential selection from a client-controlled URL path, with no in-process authorization. The description does not say path. It says body or header. The score explanation, dated 2026-10-27 with no minute, says the score was re-evaluated after the vulnerable mechanism changed and was not fixed. It says the path-based lookup was removed in commit bac9fb5, and credential selection now follows the client-supplied body model field and the x-gateway-model-name header, through a cluster-wide model-name map, still with no in-process identity check. It says subscription validation derives the model identity from the URL path, while credential injection uses the body and the header. Attack complexity moved from high to low on that reading. AV:N is the in-cluster ext_proc gRPC port. There is no CISA ADP record. datePublic is null.

Commit bac9fb55e7e6c9b184720a581054b1ffdc3f62c6 is pull request 359, merged at 2026-07-03T21:04:40Z. The author is Noy Itzikowitz, login noyitz, and the committer time is the same second. The subject is a feature: resolve ExternalModel by body model name for single-URL routing. The message says the resolver prefers the X-Gateway-Model-Name header and falls back to the body model field. Header names are case-insensitive, so that spelling and the score note's lowercase form are one header. Compare of main to this commit is behind, ahead 0, behind 63, so main contains it. This read did not diff those 63 commits for an identity check.

The repository was created 2026-03-12T13:36:58Z. pushed_at on this read is 2026-10-28T18:19:46Z, which is 1 hour, 42 minutes, 28.895 seconds before the CVE. The tip subject is a gRPC size change, commit bb1fee4547, committer 2026-10-28T18:19:45Z. The fifteen newest subjects do not name CVE-2026-117574. Releases length is 0. The only tag is v0.1.0-rc.1, commit 1080e4a56af9cbf78184723586c983177accc335, committer 2026-04-21T12:54:33Z, subject an API translation plugin. Compare of that tag to bac9fb5 is ahead, behind 0, ahead 124. The tag is an ancestor. It does not contain the July commit.

The bug is still NEW.#

Bugzilla 2494861 is product Security Response, component vulnerability, alias CVE-2026-117574. creation_time is 2026-06-30T12:16:08Z. status is NEW. resolution is empty. is_open is true. cf_last_closed is null. Severity high, priority high. last_change_time is 2026-10-27T12:46:29Z. The score note's 27 October date has no minute, so this brief does not treat that bug timestamp as the score edit.

From the bug open to the July commit is 3 days, 8 hours, 48 minutes, 32 seconds. From the commit to the CVE is 116 days, 22 hours, 57 minutes, 34.895 seconds. From the bug open to the CVE is 120 days, 7 hours, 46 minutes, 6.895 seconds. The CVE dateReserved is 2026-10-27T11:34:15.315Z. From that reserve to datePublished is 1 day, 8 hours, 27 minutes, 59.580 seconds. From the bug's last change to the CVE is 1 day, 7 hours, 15 minutes, 45.895 seconds.

The affected block names two container packages and no versions. rhoai/odh-ai-gateway-payload-processing-rhel9 has defaultStatus affected. rhoai/odh-ai-gateway-payload-processing-e2e-rhel9 has defaultStatus unaffected. Both carry cpe:/a:redhat:openshift_ai. A scanner that stops on the e2e row is reading the test image.

  1. Tag v0.1.0-rc.1, committer 12:54:33Z. It is the only tag. Compare to bac9fb5 is ahead, behind 0. The tag does not contain the July commit. Releases length is 0.
  2. Bugzilla 2494861 created at 12:16:08Z. Product Security Response. The summary still says a client-controlled URL path.
  3. PR 359 merged at 21:04:40Z, commit bac9fb55, author Noy Itzikowitz. Subject is a feature. The message prefers X-Gateway-Model-Name and falls back to the body model field. Main contains the commit.
  4. CVE reserved at 11:34:15.315Z. Bug last_change_time 12:46:29Z. Status still NEW. The score note says re-evaluated this day, no minute, because the mechanism changed and was not fixed.
  5. Repo pushed_at 18:19:46Z. Tip subject is a gRPC size change. CVE published at 20:02:14.895Z. CVSS 3.1 9.1. Red Hat word Important. No version. No ADP record.

What to do with the two clocks.#

  1. Keep the June bug, the July commit, and the October CVE as three times. The bug opened at 2026-06-30T12:16:08Z. The lookup commit merged at 2026-07-03T21:04:40Z. The CVE published at 2026-10-28T20:02:14.895Z. From the commit to the CVE is 116 days, 22 hours, 57 minutes, 34.895 seconds. The commit is the package clock. The other tenant's provider API key is the impact clock. A lookup change does not rotate that key.
  2. Leave bug 2494861 open in your tracker. Status is NEW, resolution is empty, and cf_last_closed is null. The fifteen newest commit subjects do not name the CVE. This read did not diff the 63 commits after bac9fb5.
  3. Watch the field the description names. The title says a URL path. The score note says subscription validation still reads the path, and credential injection reads the body model field and X-Gateway-Model-Name. A gateway policy that only binds the path does not cover the field that selects the key.
  4. Separate the two image names. The affected package is rhoai/odh-ai-gateway-payload-processing-rhel9. The e2e package is unaffected. Neither row has a version. v0.1.0-rc.1 is not a fix line. It predates the commit.
  5. The record does not say the flaw has been exploited. It does say an authenticated caller of one tenant can receive another tenant's provider key. Until a build is named, limit which accounts can authenticate to that gateway. The statement's boundary is that valid user credential. Provider keys a shared gateway can inject stay on the impact clock after any image that only changed the lookup.

Frequently asked

Did commit bac9fb5 close the tenant boundary?

The score note says no. It says the path-based lookup was removed in that commit and credential selection moved to the client-supplied model field, with no in-process identity check. The commit subject is a feature for single-URL routing. Bugzilla 2494861 is still NEW, with an empty resolution.

Which image name in the CVE is the affected one?

rhoai/odh-ai-gateway-payload-processing-rhel9 has defaultStatus affected. rhoai/odh-ai-gateway-payload-processing-e2e-rhel9 has defaultStatus unaffected. Neither row carries a version. The only Git tag on the upstream repo is v0.1.0-rc.1 from 21 April, and that tag does not contain bac9fb5.

Does the CVE title mean a path check is enough?

The title still says a client-controlled URL path. The description says another tenant's model identifier in the request body or header. The score note says subscription validation reads the path, and credential injection reads the body and the header. A control that only watches the path does not cover the field the description names.

Sources

  1. CVE-2026-117574 Red Hat · 2026-10-28
  2. Bugzilla 2494861 Red Hat · 2026-06-30
  3. Resolve ExternalModel by body model name opendatahub-io · 2026-07-03
  4. CVE-2026-117574 record Red Hat CNA · 2026-10-28

Related