Skip to content
AI Security Wire

Published

- 7 min read

By

GPUStack CVE-2026-102895: Loopback Trust Became an Admin Backdoor

img of GPUStack CVE-2026-102895: Loopback Trust Became an Admin Backdoor

GPUStack is an open-source GPU cluster manager for AI model serving and GPU instance provisioning. Before version 2.2.0, its authentication code had a shortcut: a request that appeared to come from 127.0.0.1 was treated as the built-in administrator. The shortcut breaks in what the advisory calls the “canonical self-hosted shape”: the service behind a reverse proxy on the same machine. CISA added the flaw, CVE-2026-102895, to its Known Exploited Vulnerabilities (KEV) catalog on November 9, 2026, with a due date of November 12.

This article explains the mechanism and what the fix changed. Facts come from the GitHub security advisory GHSA-gj9m-f8ph-pmhr, the CVE record, NVD, the CISA KEV catalog, the fix pull request, the GitHub advisories for CVE-2026-102894 and CVE-2026-103002, and a Daily Security Review article, all opened for this article. The CVSS score comes from the GitHub CNA; NVD shows it as a secondary score and has no primary score.

At a Glance

FieldDetail
CVECVE-2026-102895
ProductGPUStack (pip package gpustack)
AffectedVersions before 2.2.0
Fixed2.2.0 (the advisory lists 2.2.0 as the first patched version)
WeaknessCWE-287, CWE-290, CWE-346
CVSS 3.19.8, CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H
Score sourceGitHub CNA (the CVE assigner); NVD lists it as a secondary score and CISA-ADP adds an SSVC assessment but no score of its own
CISA KEVAdded 2026-11-09, due 2026-11-12, known ransomware use “Unknown”
ExploitationObserved in the wild against a v2.1.1 deployment, per the vendor advisory

The Mechanism

According to the advisory, get_current_user in gpustack/api/auth.py first tries to authenticate the caller from an API token. If that yields no user, and the request source is considered loopback, and force_auth_localhost is False (the default), the function looks up the first admin account and returns it. The fix pull request shows the removed lines:

   if user is None and client_ip_getter(request=request) == "127.0.0.1":
    if not server_config.force_auth_localhost:
        user = await User.first_by_field(session, "is_admin", True)

The weakness is that the request “source” comes from client_ip_getter, and the advisory describes two ways an outsider can make it report 127.0.0.1:

  1. Same-host reverse proxy. With a proxy configured as proxy_pass pointing at a loopback address, every remote request reaches GPUStack with a loopback TCP peer. The loopback branch fires with no header and no credential. The advisory says this was verified end to end against HEAD 6bee5883.
  2. Forged header. A client-supplied X-GPUStack-Real-IP: 127.0.0.1 header can cause the request source to be treated as loopback.

The advisory’s root-cause statement is that the code trusted the network source of a request (a socket peer or a client-controllable header) as an authentication signal.

What an Attacker Gets

The advisory lists the impact of unauthenticated administrative access:

  • Enumerate and modify users, models, workers and configuration.
  • Mint persistent administrator API keys (scope ["*"]) that stay valid from any source IP afterwards.
  • Read worker credentials and cloud and provider secrets.
  • Execute code on worker hosts through admin-only functionality such as POST /v2/inference-backends/from-yaml.

The CISA KEV entry repeats this list. The persistent API key matters for incident response: upgrading removes the bypass, but an admin key created earlier keeps working.

Exploitation Evidence

The advisory states the flaw “has been observed exploited in the wild against a v2.1.1 deployment (attacker-created custom backends running reverse shells)” and credits Javier Sixto (ulex.fr) for reporting that exploitation. Independent finder credits go to Nikko Enggaliano, Wei Yunzhou, and Rui Yang and Haoyu Wang of Johns Hopkins University. The CISA-ADP record in the CVE entry carries an SSVC assessment of exploitation “active”, automatable “yes”, technical impact “total”.

The CVE record and NVD repeat the maintainer statement, and the KEV entry (added 2026-11-09) does not cite independent evidence that we could see, so exploitation rests on a single source with no independent telemetry that we found. We found no evidence as of 2026-11-12 in the advisory, CVE record, NVD, KEV or Daily Security Review of who exploited it or how many instances were hit (Daily Security Review itself says the first-attack date and the compromised-server count are not stated). We could not locate the ulex.fr report. Treat the exploitation scope as unknown. Trade press such as Daily Security Review repeats the advisory’s facts.

Timeline

DateEventSource
2026-05-19Fix pull request #5346 “refactor: remove localhost auto-admin bypass” mergedGitHub PR
2026-06-18GPUStack v2.2.0 releasedGitHub releases
2026-07-31Advisory shown as published in the GPUStack repository advisory list (GitHub API published_at)GitHub API
2026-11-04CVE record published (GitHub CNA); NVD entry publishedcve.org, NVD
2026-11-09GitHub global (reviewed) advisory published; CISA adds to KEVGitHub, CISA
2026-11-12CISA KEV due dateCISA

The pull request title does not mention security, and the 2.2.0 release notes we read describe features rather than this fix. The repository advisory list dates the advisory to July 31, while the CVE record and the global GitHub advisory followed on November 4 and 9. We do not know why the CVE record came months after the fix. Our inference: operators tracking CVE feeds rather than repository advisories would have seen it late.

What the Fix Changed

In the pull request, the loopback auto-admin branch was deleted from get_current_user, and the is_local_request localhost guard on PUT /config was removed as well. --force-auth-localhost became a deprecated no-op: the documentation now says localhost callers are always authenticated, and the server logs a warning if the setting is supplied. The reset-admin-password command, which relied on localhost access, was changed to mint a short-lived admin JWT from the local secret key instead. The advisory also says client-IP and forwarding-header handling was hardened. That change is not in the pull request; client_ip_getter in gpustack/api/auth.py differs between v2.1.1 (where it returns the X-GPUStack-Real-IP header) and v2.2.0 (which adds logic around that header). We did not analyse the v2.2.0 logic.

One consequence for operators: on 2.2.0 and later, force_auth_localhost=true is no longer the control, because the flag is ignored. The workaround below applies only to vulnerable versions.

Remediation

  1. Upgrade. The advisory fixes this CVE in 2.2.0. GitHub’s advisory database also lists CVE-2026-102894 (critical, CVSS 9.1, a SAML callback that does not verify the response) and CVE-2026-103002 (high, CVSS 8.8, missing authorization on some cluster endpoints), both fixed in 2.2.2. Go to 2.2.2 or newer; the latest release we saw in the GitHub releases list was v2.3.1 (published 2026-11-11). GPUStack published a further batch of advisories on November 11 that list 2.3.1 as the fixed version, so our recommendation is 2.3.1. Caveats from the vendor advisories: 2.2.0 and 2.2.1 remain exposed to the CVE-2026-102894 SAML bypass (affected versions below 2.2.2), and CVE-2026-103002 requires authentication and a Kubernetes-provider cluster. We found no evidence as of 2026-11-12 in the releases and compare API of a 2.1.x backport of the fix.
  2. If you cannot upgrade yet (from the advisory): set force_auth_localhost=true; ensure the proxy or gateway cannot cause GPUStack to treat external traffic as loopback; strip any client-supplied X-GPUStack-Real-IP header at the edge; and keep the API off untrusted networks with a VPN or edge authentication.
  3. Assume compromise if it was exposed. KEV status and the advisory’s exploitation note both support this. Review the admin user list and API keys for entries you did not create, rotate worker and cloud-provider secrets the server stored, and inspect custom inference backends for unexpected definitions or reverse-shell commands.

Detection Guidance (our suggestions)

These are our own starting points, not vendor-published signatures:

  • Requests to administrative API routes that succeed without an Authorization header or session cookie.
  • Inbound requests carrying an X-GPUStack-Real-IP header from external clients, which should never happen in a normal deployment.
  • New admin API keys, new users, or POST /v2/inference-backends/from-yaml calls you cannot match to a change ticket.
  • Outbound connections from worker hosts shortly after a custom backend was created.

What This Means

Our assessment: the lesson is not specific to GPUStack. AI serving stacks are often deployed with a local reverse proxy for TLS, and any “localhost is trusted” rule becomes “everyone is trusted” at that point. In our view, self-hosted AI infrastructure also tends to hold valuable material (model access, cloud credentials, GPU capacity) behind a single admin role, which makes an authentication bypass a direct route to code execution on the worker hosts. Inventory GPUStack and similar model-serving managers, and check what sits in front of them.

References

Frequently Asked Questions

What is CVE-2026-102895?
An improper authentication flaw (CWE-287, CWE-290, CWE-346) in GPUStack, an open-source GPU cluster manager for AI model serving. In versions before 2.2.0, get_current_user in gpustack/api/auth.py promoted any request whose source looked like loopback to the built-in administrator when force_auth_localhost was false, which was the default. The GitHub advisory and the CVE record both score it CVSS 3.1 9.8 (AV:N/AC:L/PR:N/UI:N).
How could a remote attacker look like localhost?
The advisory describes two routes. If GPUStack sits behind a reverse proxy on the same host using proxy_pass to 127.0.0.1, every remote request arrives with a loopback TCP peer, so no header or credential is needed. Alternatively, a forged X-GPUStack-Real-IP: 127.0.0.1 header could make the request source be treated as loopback.
Is it being exploited?
Yes. CISA added CVE-2026-102895 to its Known Exploited Vulnerabilities catalog on November 9, 2026 with a due date of November 12, and the GitHub advisory states exploitation was observed in the wild against a v2.1.1 deployment, where attacker-created custom backends ran reverse shells. CISA lists known ransomware campaign use as Unknown.
What should operators do?
Upgrade to GPUStack 2.2.0 or later; separate GitHub advisories for two other GPUStack flaws (CVE-2026-102894 and CVE-2026-103002) list 2.2.2 as fixed, and our recommendation is to target 2.3.1, the latest release we saw. If you cannot upgrade immediately, the advisory's workarounds are to set force_auth_localhost=true, make sure the proxy cannot cause external traffic to be treated as loopback and strip any client-supplied X-GPUStack-Real-IP header, and keep the API off untrusted networks. Because exploitation was observed, review for unexpected admin API keys and custom inference backends.