{"resourceId":"rhoai-cve-2026-83495-trainer-v2-secretkeyref-exfiltration","versions":[{"version":"external-c5f41e41c815cd9510a1be5db7c16d8d7e829a8b230e1bde1e49c1ff28b1674e","resource":{"id":"rhoai-cve-2026-83495-trainer-v2-secretkeyref-exfiltration","title":"CVE-2026-83495: Low-impact Trainer v2 TrainJob secretKeyRef injection in OpenShift AI (Technology Preview) enables same-namespace secret exfiltration where RBAC is mismatched","organization":"Red Hat Product Security","sector":"Cross-sector technology vendor (OpenShift AI)","geography":"Global product","publishedAt":"Red Hat public date 2026-08-31; CVE page last changed 2026-11-07 (UTC)","publicationDate":"2026-08-31","eventDate":"2026-11-07","sourceName":"Red Hat Customer Portal","sourceLabel":"Vendor CVE record: Low, rhoai/odh-trainer-rhel9 Under investigation","sourceUrl":"https://access.redhat.com/security/cve/cve-2026-83495","evidenceClass":"vendor-claim","outcomeClass":"cautionary","topics":["data-security","developers-agents","governance-procurement"],"finding":"Red Hat rates CVE-2026-83495 Low because Trainer v2 is a Technology Preview feature of OpenShift AI. A user who can create TrainJobs can inject valueFrom.secretKeyRef into environment variables to read and exfiltrate namespace secrets, a privilege escalation when TrainJob RBAC is broader than Secret read RBAC. Mitigation: align TrainJob creation RBAC with Secret read permissions, or restrict secretKeyRef to an allowlist in the validating webhook. Package rhoai/odh-trainer-rhel9 is Under investigation with no advisory.","sledRelevance":"University labs and agency data-science teams often grant broad job-creation rights while restricting secret access; this flaw shows such a mismatch can expose credentials such as data-source keys within a project namespace.","evidence":"Opened Red Hat CVE page: description, Low statement citing Technology Preview, mitigation text, Bugzilla 2526350, CWE-269; record public date 2026-08-31, last changed 2026-11-07T06:02Z, package Under investigation.","architectureImplications":"Keep data-source secrets in separate namespaces or external secret stores, and avoid job-creation rights broader than secret-read rights.","governanceImplications":"Technology Preview features such as Trainer v2 should stay out of production research platforms unless formally risk-accepted.","securityPrivacyImplications":"Confidentiality risk for namespace secrets, limited to the same namespace.","caveats":"Low impact per vendor; affects a Technology Preview component; Under investigation; no CVSS score shown.","streamIds":["red-hat"],"roles":{"sales":"Sales framing: RBAC hygiene is a practical advisory service. Use this Low-impact flaw to start a conversation about least-privilege design for OpenShift AI research namespaces, especially where Technology Preview features are enabled. Keep the tone proportionate: Red Hat rates it Low and limits it to Trainer v2 Technology Preview, so do not present it as an urgent platform-wide exposure or promise fix timing (interpretation).","engineering":"Engineering framing: align job-creation and secret-read rights. Apply Red Hat's mitigation by matching TrainJob creation RBAC to Secret read permissions or allowlisting secretKeyRef in the validating webhook. Check whether Trainer v2 is enabled and keep it off where not needed. Move sensitive credentials to an external secret manager and confirm the change by attempting the injection in staging (interpretation).","delivery":"Delivery framing: run an RBAC alignment check. Script a per-namespace comparison of TrainJob create rights with Secret read rights, fix mismatches, and record decisions about enabling Technology Preview features. Acceptance is zero mismatches in production namespaces and a denied injection attempt in staging. Repeat the check whenever new research projects are onboarded, because RBAC drift is the underlying exposure, and track the package until Red Hat moves it out of Under investigation (interpretation)."},"roleDetails":{"sales":{"customerProblem":"Research and analytics namespaces may grant job-creation rights broader than secret access, enabling credential exposure.","stakeholders":["Research computing security lead","Data science platform owner"],"discoveryQuestions":["Is Trainer v2 enabled anywhere?","Do job creators also hold secret-read rights?"],"valueHypothesis":"Bounded: RBAC alignment closes the documented exposure path.","potentialEngagement":"An RBAC alignment review for OpenShift AI namespaces.","unsupportedClaims":["High severity","Cross-namespace impact","A fix timeline"]},"engineering":{"fit":"Relevant only where Trainer v2 Technology Preview is enabled.","architectureIntegration":"Trainer v2 TrainJob API environment passthrough with secretKeyRef.","prerequisites":["RBAC inventory","Record of preview-feature enablement"],"constraints":["Technology Preview component","Package Under investigation"],"security":"Align RBAC or allowlist secretKeyRef in the webhook; use an external secret store.","proofOfValue":["Attempt secretKeyRef injection in staging after mitigation and confirm denial","Run an RBAC mismatch report"]},"delivery":{"work":"Remediate RBAC mismatches for TrainJob creators.","dependencies":["RBAC tooling","Webhook configuration"],"ownership":"OpenShift AI platform team.","skillsAdoption":"Platform admins need an RBAC review procedure.","governanceCheckpoints":["Decision record on Technology Preview enablement","RBAC change approval"],"acceptanceCriteria":["Proposed: zero namespaces where TrainJob create exceeds Secret read","Proposed: staging injection test denied"],"risks":["Breaking legitimate training jobs","Overlooked namespaces"]}},"retrievedAt":"2026-11-09T22:04:02-06:00","enrichedAt":"2026-11-09T22:20:00-06:00","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Not addressed in the source.","procurementImplications":"Not material beyond standard vulnerability-response SLAs.","operatingModelImplications":"RBAC reviews should compare job-creation and secret-read permissions per namespace.","sourceVerification":{"openedUrl":"https://access.redhat.com/security/cve/cve-2026-83495","referenceExcerpt":"This issue has a Low impact as it affects Trainer v2, which is a Technology Preview feature in Red Hat OpenShift AI.","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}