What happened
CVE-2026-112047 (CWE-863, CVSS 5.4 v3.1 / 5.3 v4.0, status Deferred) is an incorrect-authorization flaw in the Kubernetes cluster AI access API: perimeter checks are project-wide rather than per-cluster, so a project member can target clusters outside their scope through the AI/agent and kubectl approval endpoints.
Why it matters
OneUptime's AI agent and kubectl-approval controls are exactly the privileged agentic operations that should be tightly scoped; the bypass lets a low-privileged project member steer agent/Kubernetes behavior on clusters they do not own, undermining the RBAC boundaries around the platform's AI automation.
Attack vector
An authenticated project member abuses project-wide permission checks to invoke out-of-scope features — AI cluster access test, agent reset, kubectl plan approval, and incident AI conversation — learning agent RBAC permissions, disconnecting agents, approving kubectl changes, or running commands on unauthorized K8s resources.
Affected systems
OneUptime 14.0.2 through <14.0.69
Mitigation
Upgrade to 14.0.69 (fix commit 264e1ba13149cbd4c80cb22f2177b405886b0452); advisory GHSA-8fph-9j9f-xxgc.