GPUStack Authentication Bypass Draws Exploit Attempts
Attempts to exploit an authentication-bypass flaw in GPUStack have been observed targeting exposed services, according to monitoring reported on November 12, 2026. The issue, tracked as CVE-2026-102895, affects GPUStack versions earlier than 2.2.0. The flaw allows a request to be treated as originating from a trusted local address when it supplies a spoofed HTTP header, potentially bypassing authentication and access controls.
GPUStack is open-source software used to manage clusters of GPUs, which are commonly deployed for artificial intelligence workloads. That role can make a GPUStack instance a sensitive part of an organization’s infrastructure: access to its management interface may expose administrative functions or resources used to operate GPU workloads. The report does not establish that any targeted environment was successfully compromised, and it does not quantify the scale of the observed activity.
How the bypass works
The vulnerable behavior concerns the X-Gpustack-Real-IP header. Headers such as Forwarded, X-Forwarded-For and X-Real-IP are often used by proxies to convey a client’s address to an application. They are useful for logging and request handling, but they cannot safely serve as proof of identity if an untrusted client can supply or alter them.
In this case, GPUStack makes an access-control decision based on the address conveyed by X-Gpustack-Real-IP. If the value indicates the loopback address 127.0.0.1, GPUStack treats the request as coming from a trusted local source and bypasses authentication. An attacker able to send a request with that header can therefore attempt to gain access without valid credentials. The underlying security concern is not limited to the header’s name: any client-controlled value used to establish trust can undermine the control meant to protect the application.
Observed requests and timeline
The reported attempts used a straightforward HTTP GET request to the root path and included X-Gpustack-Real-IP set to 127.0.0.1. The monitoring report identified 144.91.73.103 as the source IP address for those requests. It also noted that, over the preceding couple of days, the same address had looked for other AI-related endpoints, including paths associated with model listings and API requests. These observations indicate probing; they do not, by themselves, demonstrate successful authentication bypass or access to data.
GPUStack versions before 2.2.0 are affected. Version 2.2.0 was released at the end of June 2026, while the vulnerability was made public at the end of July. The CVE identifier was assigned on November 4, 2026. GPUStack’s security advisory is available under identifier GHSA-gj9m-f8ph-pmhr. Organizations should consult that advisory and their normal software-management channels to confirm the applicable remediation and deployment status.
Actions for operators
Organizations running GPUStack should first identify deployed instances and determine whether any are earlier than 2.2.0. Where an affected version is present, prioritize upgrading to a fixed release in line with the project’s advisory. Teams should also assess whether GPUStack management interfaces are reachable from untrusted networks and restrict access to authorized administrative paths where practical.
- Review application, proxy and firewall logs for requests carrying X-Gpustack-Real-IP with a loopback value, including requests to the root path and AI-related API endpoints.
- Check for activity from 144.91.73.103 as one supplied indicator, while recognizing that a single address is not a complete detection strategy.
- Ensure that trusted client-address headers are set or overwritten by a controlled proxy and cannot be accepted directly from untrusted clients.
- Review other applications for authorization decisions that rely on client-supplied forwarding headers rather than verified connection metadata.
The incident highlights a recurring trust-boundary problem for services deployed behind proxies: forwarding metadata can support routing and observability, but applications must not treat it as authentication unless its origin is reliably controlled. For GPUStack operators, version assessment, remediation and focused log review are the immediate priorities.