Critical GPUStack RCE Vulnerability CVE-2026-102895 Exposes Enterprise AI Compute Clusters
A critical unauthenticated remote code execution flaw in open-source orchestrator GPUStack allows threat actors to seize enterprise AI accelerator clusters.

By Elena Rostova, Cybersecurity
, updated 6 min read
Cybersecurity researchers have uncovered a maximum-severity vulnerability in GPUStack, a widely deployed open-source software platform used by enterprises to orchestrate, virtualise, and serve large language models across distributed GPU clusters. Tracked as CVE-2026-102895 with a Common Vulnerability Scoring System (CVSS) score of 9.8, the flaw permits unauthenticated remote threat actors to achieve arbitrary remote code execution on host server nodes.
Active exploitation has been observed in the wild, with threat actors targeting internet-accessible GPUStack management interfaces to hijack valuable Nvidia H100, A100, and RTX workstation clusters. Attackers have leveraged compromised compute clusters to exfiltrate proprietary proprietary model weights, siphon training datasets, and deploy automated cryptocurrency mining payloads.
Missing authentication on internal administrative endpoints
GPUStack allows organisations to manage heterogeneous hardware clusters, coordinating model distribution, inference scaling, and hardware resource allocation across local Apple Silicon, Linux, and Windows nodes through a unified web interface and REST API.
The vulnerability stems from improper authentication verification within the backend FastAPI route dispatchers. While public-facing user endpoints enforce token validation, several internal worker orchestration endpoints failed to verify JSON Web Tokens (JWT) or administrative API secrets when handling specific HTTP request methods.
By sending a specially crafted HTTP POST request to the unauthenticated worker registration and task scheduling endpoints, an external attacker can instruct the GPUStack coordinator to deploy a containerised worker payload. Because worker processes require root or high-privilege system access to configure GPU kernel drivers and memory mappings, the attacker's commands execute with full administrative privileges on the underlying operating system.
High-value targets: Model theft and compute hijacking
As enterprise organisations accelerate internal AI development, high-end GPU clusters have become prime targets for cybercrime syndicates and state-sponsored espionage groups. Nvidia server clusters running tens or hundreds of high-end enterprise GPUs represent millions of dollars in capital investment and house organisations' most valuable proprietary algorithms.
Threat intelligence firms reported that opportunistic threat groups began scanning global IPv4 address ranges for exposed GPUStack default ports within hours of proof-of-concept exploit code circulating on technical discussion boards.
In several documented intrusions, attackers used the remote code execution vector to access local storage volumes housing fine-tuned foundation models, copying proprietary weights to external command-and-control servers. In other instances, threat actors installed persistent backdoors and repurposed idle GPU compute cores to mine Monero and other privacy-focused digital tokens, causing extreme thermal throttling and system instability for legitimate enterprise workloads.
Emergency patching and defence recommendations
The GPUStack open-source development team, in coordination with vulnerability disclosures from security researchers, published emergency release GPUStack version 1.2.4. The patch introduces strict, universal authentication middleware that intercepts and validates credentials across all internal and external route handlers before parsing request parameters.
Security teams managing on-premises or cloud-hosted AI infrastructure should implement the following protective measures immediately:
Upgrade GPUStack instances to version 1.2.4 or later without delay.
Verify that GPUStack management consoles and worker ports (default TCP port 80 and port 10150) are never exposed directly to the public internet. Restrict access strictly to internal corporate subnets or authenticated VPN tunnels.
Inspect system logs and container runtimes on all GPU worker nodes for unauthorised Docker or Podman containers launched outside standard CI/CD deployment pipelines.
Rotate all API keys, database credentials, and cloud access secrets stored in environment variables on compromised host machines.
Frequently Asked Questions
What is GPUStack?
GPUStack is an open-source orchestration platform that enables enterprises to run, manage, and scale AI models across distributed GPU infrastructure, including Nvidia, Apple Silicon, and AMD hardware.
How does CVE-2026-102895 allow remote code execution?
A flaw in API route handling allows unauthenticated remote requests to trigger worker deployment commands, executing arbitrary shell code on the host server with administrative root privileges.
Has this vulnerability been exploited in the wild?
Yes. Security telemetry confirms active internet scanning and exploitation attempts targeting internet-facing GPUStack servers to steal AI model weights and deploy cryptocurrency mining tools.
Sources
- 1GPUStack Security Advisorygithub.com
- 2National Vulnerability Databasenvd.nist.gov