Skip to content
Log inRegister

GPUStack CVE-2026-102895: CVSS 9.8 Admin Bypass Exploited and Added to CISA KEV

Dr. Abeer Alshammari · Published 11/10/2026

IntermediateProfessionalsCISOsExecutives

Bottom line: GPUStack versions before 2.2.0 can hand full administrator access to anyone who can reach the API through a same-host reverse proxy. Exploitation has been observed and CISA has added the flaw to KEV. Upgrade, then check for attacker-created admin keys.

What happened

CVE-2026-102895 is an authentication bypass in GPUStack, a platform used to deploy and manage AI inference workloads across GPU hosts. By default the server treats any request that appears to come from the loopback address 127.0.0.1 as the built-in administrator. When GPUStack sits behind a reverse proxy on the same host, every remote request reaches it from loopback, so no credentials are needed. In embedded gateway mode a forged X-GPUStack-Real-IP header with the value 127.0.0.1 has the same effect. The flaw is scored CVSS 9.8. According to the published advisory and coverage of CISA's 9 November additions, it has been exploited in the wild, and the vendor advisory describes attackers creating inference backends that ran reverse shells. I could not open the CISA alert page itself, so the KEV listing is based on secondary coverage.

Why it matters

AI infrastructure is being deployed quickly, often by teams outside traditional security review, and it concentrates expensive GPUs, model weights, datasets and cloud credentials. An unauthenticated admin on the management plane turns that concentration into a single point of compromise.

Who is affected

Any organisation running GPUStack 2.1.1 or earlier, especially where the API is published through a reverse proxy on the same machine, which is a common deployment pattern. Installations with force_auth_localhost left at its default of false are in scope.

Technical impact

An attacker gains the built-in admin account, can create persistent admin API keys, and can register a custom inference backend from YAML, which leads to command execution on worker hosts. Persistent keys mean the attacker keeps access even after the underlying bug is fixed.

Governance impact

This is a reminder that AI platforms belong in the asset inventory, the vulnerability management scope and the third-party risk register like any other internet-facing system. Federal US agencies face a KEV remediation deadline, and regulators and auditors in other jurisdictions increasingly treat KEV listing as evidence that a risk was known and should have been prioritised.

What security teams should do

Upgrade to GPUStack 2.2.0 or later and recreate the container, since pulling a new image alone does not replace the running process. If you cannot upgrade immediately, set force_auth_localhost to true, strip any client-supplied X-GPUStack-Real-IP header at the proxy, and restrict the API to trusted networks or a VPN. After upgrading, review admin users, API keys (especially broadly scoped ones) and registered inference backends, revoke anything unexpected, rotate worker and cloud credentials, and look for unusual outbound connections from worker hosts.

What executives should know

If your teams run internal AI platforms, ask whether they are inventoried, patched and kept off the public internet. Exploitation is already happening, so this is a same-week action, and the investigation should cover credentials the platform could reach, not only the patch.

Sources checked: GitLab Advisory Database entry for CVE-2026-102895, Ostorlab technical analysis, and securityonline.info coverage of the CISA KEV additions of 9 November 2026. The CISA alert page could not be opened, so the KEV listing is reported from secondary coverage. Summarised in my own words; verify versions against the vendor advisory before acting.

Try it yourself

An interactive CyberAbeer experience for this topic is in development.

Coming soon
Back to insights