What is GPUStack?
GPUStack is an open-source cluster manager for running AI models on your own hardware. It pools GPUs across a set of worker machines, schedules inference workloads onto them, and exposes OpenAI-compatible API endpoints so applications can talk to locally hosted large language models, embedding models, and image models the same way they would talk to a hosted provider. A server node holds the control plane, the user database, and the configuration; worker nodes run the actual inference backends such as llama.cpp, vLLM, and vox-box.
That design makes a GPUStack server a concentrated target. It stores API keys, worker credentials, and in many deployments cloud-provider secrets used to scale GPU capacity. It has a documented administrative path for defining new inference backends, which is a path to running code on every worker in the pool. And it is commonly placed behind an nginx or Traefik reverse proxy on the same host so it can be reached from a browser. The project is widely deployed: its Docker Hub images had been pulled roughly 942,000 times and the repository carried about 5,978 stars as of 2026-11-05.
Overview
CVE-2026-102895 is an authentication bypass in GPUStack's API authorization layer. In versions before 2.2.0, any request that the server resolves as coming from the loopback address 127.0.0.1 is treated as the built-in administrator, with no credential of any kind, whenever the force_auth_localhost setting is false. False is the default.
Two ordinary deployment patterns make a remote request look like loopback. The first is a same-host reverse proxy: when nginx or an equivalent sits in front of GPUStack and connects to it over 127.0.0.1, every external request arrives with a loopback socket peer, so every external request is admin. No header and no credential are needed. The second applies in embedded gateway mode, where the server honours the X-GPUStack-Real-IP header: a client can simply send X-GPUStack-Real-IP: 127.0.0.1 and be promoted the same way.
From there the impact is total. An attacker can enumerate and modify users, models, and workers, mint a persistent administrator API key with the * scope that survives any later proxy change, read worker and cloud-provider secrets held by the server, and call POST /v2/inference-backends/from-yaml to define a malicious inference backend, which executes on worker hosts. The vendor advisory records that this chain was used in the wild against a v2.1.1 deployment to run reverse shells. CISA added the CVE to the KEV catalog on 2026-11-09 with a BOD 26-04 deadline of 2026-11-12.
Affected Versions
| Product | Vulnerable | Fixed |
|---|---|---|
GPUStack (pip package gpustack, Docker images) |
All versions before 2.2.0, with force_auth_localhost at its default of false |
2.2.0 |
| GPUStack, recommended target | - | 2.2.2 or later, which also fixes the separate SAML flaw CVE-2026-102894 |
Technical Details
The flawed logic lives in get_current_user in gpustack/api/auth.py. After the normal credential checks fail, the dependency falls through to a branch that compares the resolved client IP against the literal string 127.0.0.1 and, on a match, returns the built-in administrator. The resolved IP comes from client_ip_getter, which returns the X-GPUStack-Real-IP header value in embedded gateway mode and otherwise the raw socket peer address. Neither input is trustworthy from the server's position: the header is attacker-controlled, and the socket peer is whatever the local proxy happens to be.
The underlying weakness is CWE-287, improper authentication, with CWE-290 (authentication bypass by spoofing) and CWE-346 (origin validation error) as the more specific descriptions. In plain words: the software used the network location a request appeared to come from as proof of who sent it, and that appearance can be manufactured. The loopback branch also applies no restriction by HTTP method or path, so it grants the full administrative API rather than some limited local-management subset.
The attack is network-reachable, needs no privileges and no user interaction, and the first step is a single unauthenticated HTTP request. CVSS 3.1 scores it 9.8 CRITICAL (AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H). It is trivially automatable against any instance fronted by a same-host proxy.
Discovery
The flaw was reported independently by several parties, which the GPUStack advisory credits as Nikko Enggaliano, Wei Yunzhou, and Rui Yang and Haoyu Wang of Johns Hopkins University. The fix commit landed on 2026-05-19 and shipped in release 2.2.0 on 2026-06-18, with the GitHub Security Advisory following on 2026-07-31 and the CVE identifier reserved on 2026-09-29.
The exploitation report is a separate contribution: Javier Sixto of ulex.fr reported the active in-the-wild attacks, and that report is what the advisory added when the CVE was published on 2026-11-04. A public proof-of-concept has since circulated on exploit aggregators.
Exploitation Context
The vendor states the vulnerability has been observed exploited in the wild against a v2.1.1 deployment, with attackers creating custom inference backends that launched reverse shells on worker hosts. That is the concrete evidence; no named threat actor, campaign, or victim count has been published, and CISA has not flagged the entry for known ransomware use.
No exposure census from Shadowserver, Censys, or GreyNoise had been published as of 2026-11-09. The adoption figures above indicate a large installed base but not how much of it is internet-facing. The practical risk estimate is different from the usual one here, because exposure is not simply a matter of whether the port is open: an instance reachable only through a co-located reverse proxy, which is the recommended production layout, is the worst case rather than a mitigated one.
Note that the gap between the fix and the KEV listing is almost five months. Any deployment still on a pre-2.2.0 build has been exploitable for that entire window, and because step three of the chain mints a persistent admin API key, an upgrade alone does not evict an attacker who already visited.
Remediation
- Upgrade to GPUStack 2.2.2 or later. Version 2.2.0 is the minimum that fixes this CVE, but 2.2.2 also closes CVE-2026-102894, an unauthenticated SAML callback account takeover scored 9.1, which is present in the same builds.
- If the upgrade cannot be done immediately, set
force_auth_localhost=trueso that loopback requests must still authenticate. This is the single most effective stopgap. - Strip any client-supplied
X-GPUStack-Real-IPheader at the reverse proxy or gateway, and make sure the proxy does not present external traffic to GPUStack as loopback. Only accept a real-client-IP header from a trusted, authenticated gateway. - Keep the GPUStack API off untrusted networks entirely, behind a VPN or an authenticating edge, rather than relying on the application to decide who is local.
- Assume compromise on any instance that ran a pre-2.2.0 build while reachable. Audit administrator accounts and API keys and revoke anything unrecognized, since a key minted through this bypass remains valid after the upgrade. Review the list of defined inference backends for entries nobody created, and rotate worker credentials and any cloud-provider secrets stored in GPUStack.
- Review proxy and application logs for unauthenticated requests to administrative endpoints, for
X-GPUStack-Real-IPheaders arriving from outside, and for calls toPOST /v2/inference-backends/from-yaml. Look back to at least 2026-06-18, and preserve logs and system state before remediating, in line with CISA's Forensics Triage Requirements under BOD 26-04. - Federal civilian agencies must remediate or discontinue use by 2026-11-12. Given a published proof-of-concept, confirmed exploitation, and a single-request entry point, every other operator should treat that date as the outer limit.
Key Details
| Property | Value |
|---|---|
| CVE ID | CVE-2026-102895 |
| Vendor / Product | GPUStack - GPUStack |
| NVD Published | 2026-11-04 |
| NVD Last Modified | 2026-11-09 |
| CVSS 3.1 Score | 9.8 |
| CVSS 3.1 Vector | CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H |
| Severity | CRITICAL |
| CWE | CWE-287 find similar ↗ |
| CISA KEV Added | 2026-11-09 |
| CISA KEV Deadline | 2026-11-12 |
| Known Ransomware Use | No |
CVSS 3.1 Breakdown
Required Action
Timeline
| Date | Event |
|---|---|
| 2026-05-19 | Fix commit removing the loopback auto-admin path lands in the GPUStack repository |
| 2026-06-18 | GPUStack 2.2.0 released, carrying the fix |
| 2026-07-31 | GitHub Security Advisory GHSA-gj9m-f8ph-pmhr published |
| 2026-09-29 | CVE-2026-102895 reserved and the advisory updated |
| 2026-11-04 | CVE published; the advisory adds a report of in-the-wild exploitation against a v2.1.1 deployment |
| 2026-11-09 | Added to CISA Known Exploited Vulnerabilities catalog |
| 2026-11-12 | CISA BOD 26-04 remediation deadline |
References
| Resource | Type |
|---|---|
| NVD - CVE-2026-102895 | Vulnerability Database |
| CISA KEV Catalog Entry | US Government |
| GPUStack Advisory GHSA-gj9m-f8ph-pmhr - Authentication bypass via trusted loopback source | Vendor Advisory |
| Ostorlab - GPUStack Auth Bypass RCE, Exploited in the Wild | Security Research |
| Daily CyberSecurity - Attackers Exploit Critical GPUStack Authentication Bypass in the Wild to Run Reverse Shells | News |
| VulDB - CVE-2026-102895 in GPUStack | Vulnerability Database |
| CISA Alert - CISA Adds Four Known Exploited Vulnerabilities to Catalog | US Government |