Vuln in NVIDIA AIStore: Scalable Storage meets RCE - CVE-2026-65162 - a critical 10.0 vulnerability that allows remote attackers to supply unauth K8s host-namespace
retained text
A Critical 10.0 in a popular storage layer under production AI. Because the vuln is in orchestration, the exploit just looks like normal job creation. (This one was a bit overlooked, has there been other cyber or NVIDIA news lately?) CVE-2026-65162 in NVIDIA AIStore: an unauthenticated ETL request could stand up a container with the host's process namespace and full privileges, then reach the GPU node underneath. If you're running AIStore: 1. Move to the fixed 5.x line. 2. Require authentication, and keep ETL off where you don't use it. 3. Enforce admission policy that rejects privileged and host-namespace pods. No public exploitation (reported) or tooling seen (yet). Given prevalence and timelines, we won't be publishing an exploit POC. After seeing the bulletin, Stealthium reproduced the attack end to end in an isolated lab, and confirmed runtime detection. Nothing crazy fancy, it's behavioral detection that hunts for escapes and privilege escalation moments, which allows us to catch this and other attacks like it (and not be singular CVE or signature specific!)
document fetched 2026-11-19 00:24 · canonical https://reddit.com/r/cybersecurity/comments/1xzfspo/vuln_in_nvidia_aistore_scalable_storage_meets_rce