# AIStore: Scalable Storage meets RCE (CVE-2026-65162)

> Source: <https://stealthium.io/blog/aistore-cve-2026-65162-rce>
> Published: 2026-11-18 00:00:00+00:00

**Recommendation:** Upgrade AIStore to the 5.x line and require authentication on the ETL service; detect escapes by alerting on workload containers granted host namespaces or reaching outside their boundary (MITRE ATT&CK T1611).
More in the playbook "Securing agents and AI tools": <https://wpnews.pro/playbooks/agent-security>

*CVE-2026-65162 is a maximum-severity (CVSS 10.0) flaw in NVIDIA AIStore's ETL subsystem: an unauthenticated request can stand up a workload container that holds the host's process namespace and full privileges, and from there, the GPU node itself. NVIDIA addressed it in the AIStore 5.x line. Our team flagged the bug straight from the security bulletin, reproduced it in an isolated lab to confirm the impact first-hand, and showcased our escape detection at runtime.*

October 13, NVIDIA published an AIStore security bulletin with eight CVEs. Right at the top was CVE-2026-65162, scored Critical, 10.0. Remote code execution, made available by the scalable storage layer under a big share of production AI. Let’s dive in!

AIStore is NVIDIA’s elastic distributed storage stack that handles the high-throughput, read-heavy demands of artificial intelligence and machine learning (AI/ML) workflows. Specific here, its ETL feature runs user-supplied transform containers next to the data, and the vulnerability centers on how those transform jobs are admitted.

We noted this vulnerability the week the bulletin dropped, reproduced it end to end in an isolated lab to validate it really does reach the node, and confirmed runtime detection for the escape before any public exploit writeup existed.

## Why The Bug Scored A 10

Short version: when vulnerable, a legacy ETL’s job-creation path trusted the caller to describe the container it wanted, and didn’t check permission levels. By default, these requests were unauthenticated, and it was possible to ask for a container that holds the host's process view and full privileges. It’s then a short hop to reading and writing the node's filesystem. For GPU nodes, that impact extends to the whole box: other tenants' data, and the credentials on disk.

**Hard to catch, as it looks like normal orchestration:** Running containers next to data is exactly what the ETL feature is for. If you’re watching the API, you’ll just see a job being created. The container is a routine job submission, and the escape itself is ordinary filesystem and process activity, nothing malformed, so nothing that would trip behavioral or signature-based detections.

**Critically, it’s quick:** The overly-permissive (and potentially hostile) container can do its work and be gone in seconds, which sets an incredibly aggressive timeline for detection and response to try and contain an attack.

## Stealthium Detection

As Stealthium instruments the kernel beneath the workload, we can see what a process actually does on the host, details beyond what an API says it intended.

This telemetry gives us two independent opportunities to catch this and other similar attacks, without a signature tuned to this one CVE:

**At creation:** Lower-fidelity, a workload container coming up with the host's namespaces and full privileges is a standing risk on its own, and we surface it the moment it appears, from the authoritative view of what the kernel actually granted rather than a label that may or may not be present.

**During the act:** Higher-fidelity, a workload container reaching out of its own boundary into the node. That’s the escape. A legitimate transform job has got no reason to do it, and so that’s a critical alert from us, **regardless of how the hostile container came to exist.**

Because this detection looks for escape behavior, it’s not specific to AIStore: it covers this CVE and wherever the next host-namespace escape may come from.

Each alert names the container and the process, explains the technique, and maps to MITRE ATT&CK T1611 (Escape to Host).

## Security Recommendations

This is a live 10.0. We haven’t yet seen public exploitation, nor public exploitation tooling. That doesn’t mean much these days though! We won’t be sharing the full how-to.

**Top recommendations:** Move to the fixed version ASAP. Per NVIDIA's bulletin, head to the 5.x line.

**Require authentication** on the service, and keep the ETL feature off where you do not use it.

**Enforce admission policy** (Pod Security Admission or an equivalent) that rejects privileged and host-namespace pods in the namespaces that run AIStore, so a defective request is refused pre-scheduling.

Between the day a CVE is published and the day the last affected cluster is patched, the only thing standing between the flaw and the node is what is watching at runtime. If you run AIStore, and can’t upversion, and want to talk specifics behind closed doors, reach out.

As always, we believe production AI requires runtime security.

**See Stealthium in action. Book a demo.**

## Methods and Attribution

**Scope of this post:** We reproduced CVE-2026-65162 end to end in an isolated lab to confirm its impact and to build detection. We are intentionally withholding the reproduction, the crafted request, and the escape technique while fixed releases are still rolling out.

**Detection:** Keyed on the runtime behavior of a host-namespace container escape: the creation of a privileged, host-namespace workload container and, with higher fidelity, a workload container reaching into the node, mapped to MITRE ATT&CK T1611 (Escape to Host). Validated against our own reproduction.

**Vulnerability record:** CVE-2026-65162, CVSS 3.1 base score 10.0 Critical (NVIDIA CNA), CWE-250 Execution with Unnecessary Privileges. Reported by Mohamed Lemine Ahmed Jidou. References: [NVIDIA AIStore security bulletin](https://github.com/NVIDIA/product-security/tree/main/2026/5889), [NVIDIA/aistore](https://github.com/NVIDIA/aistore).
