CVE-2026-72664 is a missing authorization vulnerability in Kibana where detection rule authors can attach endpoint response actions to their rules without holding endpoint privileges. The flaw sits at the intersection of two independently-secured feature domains: detection rule authoring and endpoint response operations (host isolation, process termination, file execution). Each domain has its own privilege gates, but no check exists at the point where they compose — associating a response action with a detection rule.
If you hold detection rule write access, you can configure a rule to trigger host isolation or process termination on any matching endpoint, regardless of whether you have permissions to use those response actions directly. The action executes when the rule generates an alert, giving you endpoint-level capabilities far beyond what your role should allow.
The practical impact likely exceeds the CVSS 6.5 rating suggests. A medium-severity vulnerability reachable by anyone who can author detection rules across a large fleet — where each matching alert potentially triggers host isolation — creates aggregate blast radius that CVSS doesn't capture. The attack chain is also unusually clean: no intermediate steps, no timing dependencies, no target-side configuration required.
The likely fix is a targeted authorization check at the rule-action association endpoint — something like if (!user.hasEndpointPrivileges()) throw new Forbidden() — rather than a structural rethinking of the privilege model. This is consistent with how similar gaps have been patched in this ecosystem before.
From a defensive posture: audit who holds detection rule write privileges and treat that permission as more sensitive than it appears. Review existing detection rules for any response action attachments that shouldn't be there. Map other cross-domain interactions in Kibana — particularly where detection rules intersect with other security operations — to identify similar seams. The core question is whether your current authorization data model can even express 'detection-author without endpoint-privileges' as a coherent constraint, or if you're dependent on patches at each new seam.