Skip to main content

GPUStack Auth Bypass Is Being Exploited: Patch Self-Hosted AI Servers

CVE-2026-102895 is a critical (CVSS 9.8) authentication bypass in GPUStack, an open source tool for running AI models on your own GPUs. Versions before 2.2.0 treat requests that appear to come from localhost as the administrator. CISA added it to its exploited vulnerabilities catalog on 9 November 2026. Upgrade, then audit API keys and inference backends.

On this page

What CISA flagged and who is affected

GPUStack is an open source manager for GPU clusters. Teams use it to deploy and serve AI models on hardware they control, which makes it a common building block for self-hosted and private AI setups.

On 9 November 2026 the US Cybersecurity and Infrastructure Security Agency added a GPUStack flaw, CVE-2026-102895, to its Known Exploited Vulnerabilities catalog. Federal agencies have until 12 November to apply mitigations, and the entry is marked for forensic triage, which means agencies are expected to look for signs of compromise as well as patch.

The vendor advisory rates the bug critical, with a CVSS score of 9.8. It affects every GPUStack release before 2.2.0. The fix itself is not new: version 2.2.0 shipped on 18 June 2026 and the advisory was published on 31 July, but the CVE record only appeared on 4 November. Servers that have not been updated since the spring are the ones at risk.

How a localhost shortcut became remote admin access

The root cause is a convenience feature. Before 2.2.0, GPUStack promoted any request that seemed to come from the same machine (127.0.0.1) to the built-in administrator account, with no password or token. This was the default behaviour, controlled by a setting called force_auth_localhost that shipped switched off.

The advisory describes two ways a remote attacker can look local:

  • Same-host reverse proxy. If Nginx, Caddy, or a similar proxy runs on the same machine and forwards traffic to GPUStack on 127.0.0.1, every request from the internet arrives with a loopback address.
  • Forged header. A client can send an X-GPUStack-Real-IP header set to 127.0.0.1 and be treated as local.

Security firm Ostorlab, which analysed the bug, says the header path applies in embedded gateway mode, and that the reverse proxy layout is the main case. The Cloud Security Alliance notes that the sources disagree on how widely the header trick works, and advises assuming the broader reading until you have checked your own configuration.

Putting a reverse proxy in front of a local AI server is exactly what most guides recommend for adding HTTPS. Here, that ordinary setup removed the login entirely.

What attackers do once they are in

Admin access to GPUStack is more than a dashboard. According to the advisory, an attacker can create persistent administrator API keys and read stored worker and cloud provider credentials. By posting a custom inference backend definition to the API, they can also run code on the GPU worker machines.

This has happened in practice. The advisory records exploitation against a deployment running version 2.1.1, where attackers created custom backends that opened reverse shells on worker hosts. CISA's catalog entry lists the same outcomes: user and model changes, API key creation, secret disclosure, and code execution.

No source gives a count of exposed servers. Ostorlab cites roughly 940,000 Docker Hub pulls as a measure of popularity, not of vulnerable installs.

How to patch and check for compromise

If you run GPUStack, treat this as urgent:

  • Upgrade. Version 2.2.0 removes the loopback shortcut. The Cloud Security Alliance recommends going to 2.2.2, which also fixes a separate SAML login bypass, CVE-2026-102894, and recreating containers so the new image is actually in use.
  • If you cannot upgrade today, set force_auth_localhost to true, strip any client-supplied X-GPUStack-Real-IP header at your proxy, and keep the API off untrusted networks, for example behind a VPN.
  • Hunt for leftovers. Patching does not remove an attacker who already got in. An API key with full scope stays valid after the fix. Review admin accounts and API keys, inspect inference backends you did not create, and rotate worker tokens and cloud credentials that GPUStack held.

The wider lesson applies to any self-hosted AI tool: running models on your own hardware keeps your data away from third parties, but only if the server's own front door is locked. A network address is not proof of identity, and the advisory's guidance is blunt on that point: do not derive authentication from where a request appears to come from.

CuriousLM runs supported AI models locally on your device. Try CuriousLM.