CVE-2026-102895: GPUStack Loopback Admin Hits CISA KEV - Floor 2.2.0

CVE-2026-102895 is an unauthenticated admin bypass in GPUStack before 2.2.0. GPUStack is an open-source GPU cluster manager for model serving. GitHub's CNA scores it 9.8 CRITICAL (CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H). CWE 287 / 290 / 346. Floor: gpustack 2.2.0. CISA added it to KEV on 9 November 2026, due 12 November. knownRansomwareCampaignUse Unknown. forensicTriage Yes.[1][2][3][4]
This is not a cPanel, LiteSpeed, or WordPress CVE. If the box only runs shared hosting, stop here. If you put GPUStack behind nginx/Caddy on a GPU VPS or bare-metal node - the shape the advisory calls canonical - this is the KEV that was due three days ago.
What it is
Before 2.2.0, get_current_user in gpustack/api/auth.py promotes any request it thinks came from loopback to the built-in administrator whenever force_auth_localhost is False (the default). The “source” comes from client_ip_getter. Two ways that reports 127.0.0.1 with no credentials:[1][2]
- Same-host reverse proxy.
proxy_pass http://127.0.0.1:<port>makes every remote request arrive with TCP peer127.0.0.1. No header required. The advisory says this was verified end-to-end against git6bee5883. - Forged header.
X-GPUStack-Real-IP: 127.0.0.1can be treated as loopback.
That is full admin over the network: users, models, workers, config; mint persistent admin API keys with scope ["*"] that keep working from any source IP after you upgrade; read worker credentials and cloud/provider secrets. Admin-only POST /v2/inference-backends/from-yaml is the code-execution path the advisory names. They observed it exploited in the wild on a v2.1.1 install: attacker-created custom backends running reverse shells.[1][2]
How the engine bug reads - the vendor's words, not a PoC from us. Authentication was derived from where the request appeared to come from (socket peer or a client-controlled header). A same-host proxy makes the peer always loopback, so “localhost is trusted” is false. We did not send that header at anyone's API.[1]
Fix commit: 53f42a6a553542974396d16a4ff6b12bcbedb48d. PR #5346. Release tag v2.2.0 (published 18 June 2026 - the patch is months older than KEV).[5][6]
How serious is it?
On paper: network, no privileges, low complexity, high C/I/A. That 9.8 is GitHub's CNA (security-advisories@github.com). NVD lists it as Secondary. The NVD JSON we pulled had no NIST primary score. Do not invent one.[3]
CISA on NVD (SSVC 2.0.3, Coordinator, timestamp 2026-11-05): Exploitation active, Automatable yes, Technical Impact total.[3]
KEV: catalog we pulled this morning is 2026.11.13, 1,781 entries, dateReleased 2026-11-13T20:24:18.3008Z. Row: GPUStack Improper Authentication Vulnerability, added 2026-11-09, due 2026-11-12. Ransomware field Unknown. Forensic triage Yes. Notes call it an open-source component that other products might embed.[4]
Vendor credit: Nikko Enggaliano; Wei Yunzhou; Rui Yang & Haoyu Wang (Johns Hopkins). In-the-wild report: Javier Sixto (ulex.fr). That is their attribution, not a count we measured.[1]
Who is affected
In scope from the pages we fetched:
- pip package
gpustack< 2.2.0. NVD CPEcpe:2.3:a:gpustack:gpustack:*versionEndExcluding 2.2.0.[2][3] - The canonical self-hosted shape: GPUStack behind a same-host reverse proxy to
127.0.0.1.[1] - Any install that accepts a client-supplied
X-GPUStack-Real-IPheader, proxy or not.[1] - The exploited example in the advisory: v2.1.1.[1]
Out of scope unless you prove otherwise:
- Shared hosting that never installed GPUStack. This ID does not land via
upcporlsup.sh. - GPUStack 2.2.0 or later for this CVE. PyPI and GitHub current as of this morning: 2.3.1 (released 11 November 2026). That is newer than the floor; take it. 2.3.1 also published a separate batch of GitHub security advisories the same day - different bugs, some still without CVE IDs. This post is the KEV loopback one.[7]
We did not inventory every Helm chart / Docker tag in the wild. If the running process is still < 2.2.0, you are in the CPE range regardless of how you installed it.
What to do now
- See which GPUStack is actually running. A GitHub release on disk is not the process.
pip show gpustack
python3 -c "import importlib.metadata as m; print(m.version('gpustack'))"
# docker / compose, if that is how you ship it:
docker ps --format '{{.Names}} {{.Image}} {{.Status}}'
Want 2.2.0 or higher. 2.3.1 is what PyPI returned this morning.
- Upgrade, then make the old process die. GHSA: upgrade to 2.2.0 or later. pip:
pip install -U 'gpustack>=2.2.0'
# or pin current:
pip install -U gpustack==2.3.1
If it runs in a container, recreate the container (new image and new process). A docker pull that leaves the old PID up is not a patch. That recreate step is operator practice, not a sentence in the GHSA.
- Until the new binary is up, the vendor workarounds are:[1]
force_auth_localhost=true(auth even for loopback)- strip any client-supplied
X-GPUStack-Real-IPat the proxy - keep the API off untrusted networks (VPN / edge auth)
Those are mitigations, not the fix.
-
Treat keys as burned. The advisory is explicit: admin keys minted through this path have scope
["*"]and remain valid from any source IP after the upgrade. Rotate worker credentials and cloud/provider secrets too. Review custom inference backends - that is the reverse-shell path they named on 2.1.1.[1] -
If this host also runs a panel, patch GPUStack anyway. cPanel/Plesk updates do not bump
gpustack.
Unsupported / stale versions
Anything below 2.2.0 is the affected range. There is no “only 2.1.1” carve-out in the CPE. 2.1.1 is the exploited example, not the only NVR.[2][3]
We are not publishing a supported-branch table for 1.x. If you are still on 1.x, you are below 2.2.0.
Compromise signs
Vendor-observed on 2.1.1: custom inference backends running reverse shells.[1] We have no public file names or IPs from CISA or the GHSA. Look for:
- admin API keys you did not create, especially scope
* - extra users / workers / models
- YAML-defined inference backends you did not add
- unexpected outbound shells from GPU worker processes
- cloud keys used from IPs that are not yours
Keep logs before you tidy. We are not publishing exploit steps.
FAQ
Is there a patch? Yes. 2.2.0 (tag 18 June 2026). Take 2.3.1 if you are updating today; that is current on PyPI/GitHub as of this fetch.[1][6][7]
Is this on CISA KEV? Yes. Added 9 Nov, due 12 Nov, catalog 2026.11.13 (1,781). Ransomware Unknown.[4]
Does a WordPress / cPanel / LiteSpeed update fix it? No.
Will upgrading revoke attacker keys? The advisory says those keys stay valid from any IP. Rotate them.[1]
Is 9.8 an NVD primary score? No. It is the GitHub CNA score, Secondary on NVD. NIST primary was not in the JSON we pulled.[3]
Has it been exploited? Vendor: yes, v2.1.1, reverse shells via custom backends. CISA SSVC: active. KEV listing is CISA's exploitation bar. We did not independently confirm volume.[1][3][4]
Honesty
- We did not exploit this.
- 9.8 is GitHub's vector. NVD had no primary CVSS in the record we fetched 15 Nov 2026 (
lastModified2026-11-10T05:17:40.970).[3] - “Canonical reverse-proxy shape” and the 2.1.1 reverse-shell sentence are vendor advisory text, not a lab we ran.[1]
- Patch shipped in June; KEV landed in November. Late is still late.
- Other GPUStack GHSAs dated 11 Nov 2026 (some High/Critical, several without CVE IDs) are not this CVE. Floor for this ID remains 2.2.0; 2.3.1 is the current tree we saw.
- Blogs that say “must be 2.2.2” were not used. GHSA patched-version is 2.2.0.
Sources
[1] https://github.com/gpustack/gpustack/security/advisories/GHSA-gj9m-f8ph-pmhr - vendor GHSA, published 31 Jul 2026; floor 2.2.0; in-the-wild 2.1.1
[2] https://github.com/advisories/GHSA-gj9m-f8ph-pmhr - GitHub Advisory Database, published 9 Nov 2026 22:15Z; CVSS 9.8; pip < 2.2.0
[3] https://services.nvd.nist.gov/rest/json/cves/2.0?cveId=CVE-2026-102895 - NVD 2.0 JSON, published 4 Nov 2026, analyzed, lastModified 10 Nov 2026; SSVC active/yes/total
[4] https://www.cisa.gov/sites/default/files/feeds/known_exploited_vulnerabilities.json - catalogVersion 2026.11.13, 1,781 entries; CVE-2026-102895 added 9 Nov, due 12 Nov
[5] https://github.com/gpustack/gpustack/commit/53f42a6a553542974396d16a4ff6b12bcbedb48d - fix commit; PR https://github.com/gpustack/gpustack/pull/5346
[6] https://github.com/gpustack/gpustack/releases/tag/v2.2.0 - tag published 18 Jun 2026
[7] https://pypi.org/pypi/gpustack/json - version 2.3.1 this morning; GitHub release v2.3.1 published 11 Nov 2026