On this page (7 sections)

GPUStack, an open-source GPU cluster manager used for AI model serving, has an improper authentication flaw tracked as CVE-2026-102895. According to the CVE description, in versions before 2.2.0 a request that appears to come from the loopback address can be treated as the built-in administrator, and exploitation was observed in the wild against a 2.1.1 deployment. The CISA Known Exploited Vulnerabilities (KEV) catalog added the flaw on 2026-11-09. The NVD entry describes the vulnerability and the GitHub advisory is among the references in the CVE record.

If you run GPUStack anywhere that is reachable from a network you do not fully control, or behind a reverse proxy on the same host, treat this as urgent. The fix is to upgrade to 2.2.0.

What happened

The CVE record was published on 2026-11-04. On 2026-11-09 CISA added it to its KEV catalog, which tracks vulnerabilities with evidence of active exploitation. The KEV entry sets a due date of 2026-11-12 for U.S. federal agencies, and its ransomware campaign use field reads "Unknown".

Coverage so far is thin. The research behind this article rests on the NVD and CVE records and only one independent publisher. Who is exploiting the flaw, how many servers are affected, and what attackers did after getting in have not been published in the sources we reviewed, so we do not speculate about them.

How the flaw works

According to the CVE description, the problem sits in the function get_current_user in gpustack/api/auth.py. When the setting force_auth_localhost was false, which was the default, the code treated a request whose source was loopback as the built-in administrator. That shortcut is convenient for a tool used on a single machine, but it fails when the apparent source of a request can be influenced by someone else.

The CVE description says a remote request could appear to originate from 127.0.0.1 in two ways:

  • The client_ip_getter logic accepted the peer address of a reverse proxy on the same host, so traffic forwarded by a local proxy looks like local traffic.
  • The same logic accepted a client-supplied X-GPUStack-Real-IP value of loopback, so a remote client could claim to be local.

This is a classic trust-boundary mistake: identity was derived from a network detail that the caller could control or that a proxy flattened. The record lists CWE-287, CWE-290 and CWE-346 for it.

What an attacker gets

Because the request is treated as the administrator, the impact is broad. The CVE description lists:

  • Unauthenticated administrative access, including user and model changes.
  • Creation of persistent administrator API keys.
  • Disclosure of worker and cloud-provider secrets.
  • Code execution through POST /v2/inference-backends/from-yaml.

As general reasoning rather than a sourced fact: any administrator API key created before you patch would remain valid after the upgrade unless you remove it, and any secret that was disclosed stays disclosed. That is why cleanup matters as much as the upgrade.

Who is exposed

Per the CVE description, versions before 2.2.0 with the default force_auth_localhost value (false) are in scope. The following situations are the most worrying, based on the two spoofing paths described above:

  • The GPUStack server is reachable from the internet or a shared network.
  • A reverse proxy (for TLS or routing) runs on the same host as GPUStack, since its peer address can look like loopback.
  • The server is reachable by users or systems you do not fully trust, such as other teams, tenants or partner networks.

A server reachable only on a private network has a smaller audience, but the sources do not say it is safe. If someone on that network can reach the service, or pass traffic through a same-host proxy, the same two paths may apply. Patch regardless.

How to check your exposure

First find out which version you run and how it is installed. The research does not document GPUStack's installation methods, so the commands below are general, read-only checks that assume a Python package or Docker install; adjust them to your setup.

If GPUStack was installed as a Python package, this shows the installed version. Look for a version lower than 2.2.0:

Check the installed GPUStack version (pip-based install)Shell
pip show gpustack

If you run it in a container, this lists local images with their tags so you can see which version is deployed. Look for tags older than 2.2.0:

List local container images that mention gpustack (Docker)Shell
docker images | grep -i gpustack

Then review who can reach the service. Check firewall rules, security groups and reverse proxy configuration to confirm whether the GPUStack port is open beyond the hosts that need it.

What to do now

  1. Upgrade GPUStack to 2.2.0 or later, following the v2.2.0 release page.
  2. Until you have patched, restrict network access to the GPUStack server to the hosts and users that need it, using firewall rules or a VPN. This is general hardening, not a vendor-documented workaround.
  3. Review administrator users and API keys in GPUStack and remove any you do not recognize, since the CVE description lists persistent admin API-key creation.
  4. Rotate worker credentials and any cloud-provider secrets stored in or reachable from GPUStack, since the CVE description lists their disclosure.
  5. Search your logs for requests to /v2/inference-backends/from-yaml and, if your proxy records headers, for the X-GPUStack-Real-IP header. Investigate anything you cannot explain.
  6. Inspect model, backend and worker definitions for entries you did not create, and check hosts for unexpected processes or outbound connections.
  7. If you find signs of compromise, treat the host and its credentials as breached and follow a full incident process.
Caution. Upgrading restarts the GPUStack service and can interrupt running inference workloads. Schedule a window and back up your configuration first. Do not skip credential rotation after patching: the upgrade closes the bypass but does not by itself remove admin keys created earlier or undo disclosed secrets.
Tip. The CVE record lists a pull request and a commit as references, if you want to review the code change. No workaround other than upgrading is listed in the sources we reviewed, so do not rely on settings changes alone.

Detection and monitoring

Because exploitation is confirmed, look backward as well as patching forward. The checks here are general and the research does not document GPUStack's own logging, so adapt them to what your proxy and host actually record. This grep, run on the proxy or GPUStack host, counts lines in a log file that mention the sensitive endpoint; replace the path with your real log location:

Count log lines that mention the YAML backend endpoint (Linux)Shell
grep -c "inference-backends/from-yaml" /path/to/your/access.log

A non-zero count is not proof of compromise, since administrators may use the endpoint legitimately, but unexpected callers or times deserve a closer look. A zero count is not proof of safety either: the search misses anything if logs are not retained, rotated away, or do not record the request path. The same applies to the X-GPUStack-Real-IP header, which only appears if your proxy logs request headers.

For the longer term, keep management interfaces for AI and GPU infrastructure off the public internet, and put them behind a VPN or an authenticating proxy. Webforta Scan can show what your servers present to the internet, and Webforta Monitor can alert you to unexpected certificate, DNS or header changes. Neither replaces patching this flaw.

For background on filtering traffic in front of an application, see what a WAF is, and for broader basics see our small business website security guide. Note that a filter in front of a service is not a substitute for the upgrade.

Timeline

Key dates from the CVE and KEV records
DateEvent
2026-11-04CVE-2026-102895 published; the CVE description says the issue is fixed in 2.2.0
2026-11-09Added to the CISA KEV catalog
2026-11-12CISA KEV due date for U.S. federal agencies
Exploitation was observed against a 2.1.1 deployment, according to the CVE description. The sources we reviewed do not say who the attackers are or how widely the flaw is being used, and they give no release date for 2.2.0.

Sources

This article is based on the public sources below. Details can change as vendors and researchers publish more; check the original advisories before acting on version numbers.

  1. GPUStack Improper Authentication Vulnerability (CVE-2026-102895) is being actively exploitednvd.nist.gov ·
  • GPUStack
  • CVE-2026-102895
  • CISA KEV
  • authentication bypass
  • AI infrastructure