GPUStack CVE-2026-102895: Attackers Are Owning the AI Control Plane
GPUStack is an open-source GPU cluster manager for serving AI models. CVE-2026-102895 is a CVSS 9.8 improper-authentication flaw that lets a remote, unauthenticated attacker make requests look local, bypass authentication, become administrator, read worker and cloud secrets, and run code. It's exploited in the wild and on CISA KEV. The fix is 2.2.0 — but assume pre-patch compromise.
Executive summary. CVE-2026-102895 is a CVSS 9.8 improper-authentication flaw in GPUStack, an open-source GPU-cluster manager for serving AI models. A remote, unauthenticated attacker can make requests appear to originate from the local machine, bypass authentication, and gain administrator privileges. From there: create persistent admin API keys, read worker and cloud-provider secrets, change users and models, and execute code on the cluster. It is exploited in the wild (reverse shells observed against a 2.1.1 deployment) and was added to CISA KEV with a 12 November 2026 deadline. The fix is GPUStack 2.2.0 — but because vulnerable instances may already be compromised, patching alone isn't enough: rotate everything the cluster could reach.
What it is
| Field | Detail |
|---|---|
| CVE | CVE-2026-102895 |
| Product | GPUStack — open-source GPU cluster manager for AI serving |
| Weakness | Improper authentication (requests forged to look local) |
| Affected | GPUStack before 2.2.0 |
| Severity | CVSS 9.8 (critical) |
| Status | Exploited in the wild; CISA KEV (deadline 12 Nov 2026) |
| Fix | GPUStack 2.2.0+ |
How it works
Remote request crafted to look like it came from localhost
│ (improper trust in "local" origin)
▼
Authentication bypassed ──► administrator
│
├─► create persistent admin API keys
├─► read worker + cloud-provider secrets
├─► change users and served models
▼
Code execution on the AI serving clusterThe root problem is a classic trust-the-source mistake: GPUStack treated requests that appeared local as privileged. Forge that signal and you are admin without a credential. Admin on GPUStack is admin on the model-serving plane: you can mint durable API keys for persistence, exfiltrate the secrets workers use to reach storage and cloud APIs, and run code across the fleet.
Why it needs industry attention
This is the story of 2026's new attack surface: AI control planes. Model-serving managers, inference gateways and MLOps schedulers now hold GPU capacity, proprietary models, and cloud credentials — and many were built fast, with weak authentication. The Cloud Security Alliance flagged this case as evidence that AI-serving control planes are being actively exploited. Treat your MLOps stack with the same rigor as your identity provider, because functionally it now is one.
Fix and mitigation
- Upgrade to GPUStack 2.2.0+ immediately.
- Assume pre-patch compromise. Rotate all GPUStack API keys, and every cloud-provider secret a worker could read.
- Review admin accounts and API keys for anything you didn't create — the persistence mechanism is a planted admin key.
- Network-segment the control plane. It should not be internet-reachable; put it behind VPN/zero-trust and restrict management interfaces.
How to detect it
- New admin API keys or admin users you can't account for.
- Requests claiming a local origin from non-local network paths (proxy/X-Forwarded spoofing patterns).
- Workers making unusual outbound connections or spawning shells (reverse shells were the observed payload).
- Secret access spikes against the cloud credentials the cluster holds.
Conclusion
CVE-2026-102895 shows that "it's just infra for our models" is a dangerous way to think. An AI serving cluster is a credential vault and a code-execution platform. Patch GPUStack to 2.2.0, rotate keys and cloud secrets on the assumption you were already hit, and move the control plane off the open internet.
Do this now: upgrade GPUStack to 2.2.0, rotate all API keys and cloud-provider secrets the cluster can reach, audit admin accounts/keys for unauthorized entries, and segment the control plane away from the internet.
Related: our research on prompt injection and AI agents and the LiteLLM sandbox-escape analysis.
Sources: Security Affairs, Cloud Security Alliance, eSecurity Planet, CISA KEV. Some abuse scenarios are informed inferences; confirm against the GPUStack advisory.
Related research
When the AI Guardrail Runs Your Code: LiteLLM's Sandbox Escape (CVE-2026-40217)
LiteLLM, a popular gateway in front of LLMs, ran user-supplied Python inside a hand-built sandbox to test guardrails. Attackers escape it with bytecode tricks, get code execution as root, and turn a limited API key into full admin over the AI gateway. It's exploited and in CISA KEV. Here's how to secure your AI plumbing.
When the AI Pentester Works for the Attacker: Cairn, CL-STA-1238 and Autonomous Intrusions
Palo Alto Networks' Unit 42 reports that a China-nexus operator ran a modified copy of Cairn, an open-source AI penetration-testing framework, against 19 organisations in two weeks, with confirmed access at several. The techniques were familiar; the speed and scale for a tiny team were not. Here's what happened, what it means, and the defensive priorities that actually matter.
MCP Security: How to Secure Model Context Protocol Servers
Model Context Protocol connects AI assistants to real systems: files, tickets, databases, cloud APIs. That makes every MCP server a privileged integration. The risks that matter, how to harden servers and clients, and what to log.