GPUStack Flaw Grants Remote Admin Access and Code Execution
A critical authentication bypass in GPUStack allowed remote attackers to gain admin privileges and execute code on workers.
TL;DR
- A critical vulnerability in GPUStack allowed unauthenticated remote attackers to gain full administrative access and execute arbitrary code on workers [^1, ^5].
- Active exploitation hit a version 2.1.1 cluster, prompting CISA to add the flaw to its Known Exploited Vulnerabilities catalog [^1, ^5].
Background
GPUStack is an open-source GPU cluster manager used to serve high-performance AI models like vLLM and SGLang and provision SSH-accessible GPU instances [^1, ^2]. In AI infrastructure, cluster managers coordinate compute resources, store API keys, and manage model inference backends across node networks. If an attacker bypasses authentication on a management platform, they gain direct control over cluster assets, sensitive API secrets, and worker node host environments.
What happened
In GPUStack versions prior to 2.2.0, the get_current_user function in gpustack/api/auth.py automatically promoted unauthenticated incoming requests to the built-in administrator account if the request source was treated as loopback (127.0.0.1) and force_auth_localhost was false, which was the default setting [^1, ^5]. The system's client_ip_getter classified incoming traffic as loopback either when GPUStack ran behind a same-host reverse proxy with proxy_pass http://127.0.0.1:<port> or when a request included a client-supplied X-GPUStack-Real-IP: 127.0.0.1 header [^1, ^5].
Tracked as CVE-2026-102895 with a CVSS 3.1 base score of 9.8, the vulnerability granted unauthenticated network attackers complete administrative access [^1, ^5]. Attackers could enumerate and modify users, models, workers, and system configurations, mint persistent admin API keys scoped to ["*"], read worker credentials and cloud provider secrets, and execute arbitrary code on worker hosts via the POST /v2/inference-backends/from-yaml endpoint [^1, ^5]. Active in-the-wild exploitation was reported by Javier Sixto (ulex.fr) against a version 2.1.1 deployment where attackers created custom backends running reverse shells [^1]. The vulnerability was independently reported by Nikko Enggaliano, Wei Yunzhou, and Rui Yang & Haoyu Wang of Johns Hopkins University [^1].
Maintainers patched the issue in version 2.2.0 by removing the localhost auto-admin path entirely, ensuring all incoming requests require valid authentication regardless of origin [^1, ^3, ^5]. The fix also deprecated force_auth_localhost as a no-op flag and updated the command-line admin password reset utility to authenticate using a locally signed JWT [^2, ^3]. On November 9, 2026, CISA added CVE-2026-102895 to its Known Exploited Vulnerabilities catalog, requiring US federal agencies to apply vendor mitigations or discontinue using the software by November 12, 2026 [^2].
Why it matters
Relying on network source attributes like socket peer addresses or request headers for authentication introduces severe security flaws when microservices operate behind reverse proxies [^1, ^5]. Because standard reverse proxy setups forward external web requests over local loopback connections, assuming traffic from 127.0.0.1 belongs to a trusted local admin exposes administrative API endpoints to arbitrary remote clients [^1].
For teams operating AI infrastructure, control plane compromise risks exposure of worker credentials and cloud provider keys [^1, ^5]. The observed exploitation of backend configuration endpoints to launch reverse shells shows how easily authentication bypasses translate into full execution privileges across cluster worker nodes [^1].
Practical example
Say Elena manages an AI inference cluster using a GPUStack release older than 2.2.0 deployed behind a same-host reverse proxy [^1, ^5].
An external attacker sends an unauthenticated HTTP request to the public reverse proxy [^1, ^5]. When the proxy forwards the connection to GPUStack over 127.0.0.1, client_ip_getter reports the request source as loopback [^1, ^5]. Because force_auth_localhost defaults to false, get_current_user automatically assigns the request full built-in administrator privileges [^1, ^5].
Using admin access, the attacker submits a YAML payload to POST /v2/inference-backends/from-yaml, deploying a custom backend that spawns a reverse shell on a worker host [^1, ^5]. The attacker also generates a persistent admin API key scoped to ["*"] to maintain long-term access [^1].
To secure the environment, Elena upgrades GPUStack to version 2.2.0, eliminating the localhost auto-admin shortcut and requiring authentication for all requests [^1, ^3, ^5].