TL;DR: Update GPUStack to version 2.2.0 immediately. If immediate patching is not possible, follow CISA BOD 26-04 guidance and restrict external network access to GPUStack administrative interfaces pending remediation.
What Is the Vulnerability?
CVE-2026-102895 is a critical improper authentication vulnerability (CVSS 3.1 score: 9.8, Severity: CRITICAL) affecting GPUStack, an open-source GPU cluster manager used for AI model serving and GPU instance provisioning. The vulnerability is tracked under three weakness identifiers: CWE-287 (Improper Authentication), CWE-290 (Authentication Bypass by Spoofing), and CWE-346 (Origin Validation Error).
The root cause lies in the get_current_user function within gpustack/api/auth.py. When the configuration option force_auth_localhost was set to false — which was the default configuration — the application treated any request that appeared to originate from the loopback address (127.0.0.1) as belonging to the built-in administrator account, bypassing all authentication checks entirely.
The critical enabler is how GPUStack resolved the client IP address. The client_ip_getter component accepted:
- A same-host reverse proxy peer address resolving to loopback, or
- A client-supplied
X-GPUStack-Real-IPheader containing a loopback value (e.g.,127.0.0.1).
Because the application trusted this header without validation, a fully remote, unauthenticated attacker could craft a request with a spoofed X-GPUStack-Real-IP: 127.0.0.1 header and be treated as the built-in administrator. No credentials, tokens, or prior access were required.
Who and What Is Affected?
All deployments of GPUStack prior to version 2.2.0 are affected when running with the default configuration (force_auth_localhost: false). The NVD advisory specifically references confirmed exploitation against a GPUStack 2.1.1 deployment.
Organizations most at risk include:
- AI/ML teams running GPUStack to serve large language models or other AI workloads on GPU clusters.
- Cloud and on-premises operators who have exposed GPUStack's API or management interface to the internet or untrusted networks.
- Enterprises using GPUStack for GPU instance provisioning that store cloud-provider credentials or worker secrets within the platform.
The CVSS vector AV:N/AC:L/PR:N/UI:N confirms the vulnerability is remotely exploitable, requires low attack complexity, needs no privileges, and requires no user interaction — representing the highest risk profile possible for network-exposed services.
What Is the Potential Impact?
Successful exploitation grants an attacker the equivalent of built-in administrator access to the GPUStack platform. According to the NVD description and the CISA KEV catalog entry, confirmed and potential impacts include:
- Unauthenticated administrative access — full control over the GPUStack management plane without any credentials.
- User and model manipulation — the ability to create, modify, or delete user accounts and AI model configurations.
- Persistent backdoor creation — an attacker can create persistent administrator-level API keys, maintaining access even after a password reset.
- Secrets disclosure — exposure of worker credentials and cloud-provider secrets stored within the platform (e.g., AWS, Azure, or GCP API keys tied to GPU provisioning).
- Remote code execution — the endpoint
POST /v2/inference-backends/from-yamlcan be abused by an authenticated (or in this case, spoofed-authenticated) caller to execute code on the underlying system.
This vulnerability has been exploited in the wild, as confirmed by CISA's addition of CVE-2026-102895 to the Known Exploited Vulnerabilities catalog on 2026-11-09. Any GPUStack instance exposed to untrusted networks should be treated as potentially compromised if running a version prior to 2.2.0.
What Should You Do Right Now?
Given active exploitation confirmed in the wild, response should be treated as urgent. Follow these steps:
- Patch immediately: Upgrade all GPUStack deployments to version 2.2.0 or later, which resolves this vulnerability. Refer to the official GPUStack release notes and vendor advisory for upgrade instructions.
- Assume compromise if unpatched: Any GPUStack instance running a version prior to 2.2.0 that was reachable from an untrusted network should be treated as potentially compromised. Conduct forensic triage before returning the system to production, in line with CISA's Forensics Triage Requirements referenced in the CISA KEV required action.
- Rotate all secrets: Revoke and rotate all API keys, cloud-provider credentials, and worker secrets stored in or accessible via GPUStack. Assume these may have been exfiltrated.
- Audit administrator accounts and API keys: Review all administrator accounts and API keys for signs of unauthorized additions. Remove any that cannot be verified as legitimate.
- Network segmentation (interim): If patching cannot occur immediately, restrict access to GPUStack's API and management interface to trusted IP ranges only. Remove any public internet exposure. Per CISA BOD 26-04 guidance, if mitigations are unavailable, consider discontinuing use of the product until remediation is possible.
- Review reverse proxy configuration: Ensure any upstream reverse proxy is correctly configured to strip or override client-supplied IP headers such as
X-GPUStack-Real-IPbefore forwarding requests to backend services. - Monitor CISA and vendor channels: Stay current with CISA KEV updates (cisa.gov/known-exploited-vulnerabilities-catalog) and the official GPUStack repository for further guidance.
How Continuous External Testing Surfaces This Exposure
One of the most dangerous characteristics of CVE-2026-102895 is that it is a default-configuration vulnerability on a network-exposed service. This is precisely the category of risk that internal patch tracking and change management processes frequently miss — the service may have been deployed correctly per internal standards and yet remain vulnerable simply by virtue of its default settings.
Continuous external attack surface monitoring and network/host testing addresses this gap in several ways:
- Service discovery: External scanning identifies exposed management APIs and cluster management interfaces that may not be known to the security team, particularly in dynamic AI/ML infrastructure where new services are spun up rapidly.
- Version fingerprinting: Identifying the GPUStack version exposed externally allows immediate correlation against known-vulnerable version ranges (pre-2.2.0) without waiting for internal asset inventory to be updated.
- Authentication posture checks: Testing whether administrative endpoints accept requests without valid credentials — the precise class of failure here — is a standard check in external host and network assessments.
- Header injection surface analysis: External testing can identify whether services trust client-supplied headers for IP resolution, a pattern indicative of authentication bypass risk.
- Rapid KEV correlation: When CISA adds a CVE to the KEV catalog, continuous external testing platforms can immediately cross-reference discovered assets against the catalog, flagging affected deployments within hours rather than weeks.
For AI infrastructure teams, where GPU clusters and model-serving APIs are expanding rapidly and often outside traditional IT governance, continuous external visibility is not optional — it is a core control.
Sources & further reading: NVD CVE-2026-102895 and the CISA KEV catalog are the authority on affected and fixed versions. Vendor/independent analysis: github.com, github.com, github.com, github.com. This advisory contains no exploit code.
Frequently asked questions
Is this vulnerability being actively exploited right now?
Yes. CVE-2026-102895 has been confirmed exploited in the wild. CISA added it to the Known Exploited Vulnerabilities (KEV) catalog on 2026-11-09, and the NVD description explicitly references observed exploitation against a GPUStack 2.1.1 deployment. Any unpatched, network-exposed GPUStack instance should be treated as a high-priority risk.
Does this only affect GPUStack instances exposed to the internet?
Internet-exposed instances are at the highest risk, but the vulnerability is exploitable by any network-adjacent attacker who can send HTTP requests to the GPUStack API. Internal network exposure to untrusted segments (e.g., shared cloud VPCs, contractor networks) should also be evaluated as a threat vector.
If I cannot patch immediately, is there a workaround?
The most effective interim mitigation is to restrict network access to GPUStack's API and management interface to trusted hosts only, and to ensure any upstream reverse proxy strips client-supplied IP headers such as X-GPUStack-Real-IP. CISA's required action under BOD 26-04 also states that if mitigations are unavailable, organizations should consider discontinuing use of the product. These are interim measures only — patching to version 2.2.0 is the definitive fix.
Should I assume my GPUStack instance was compromised if it was running a vulnerable version?
If your GPUStack instance was running a version prior to 2.2.0 and was reachable from any untrusted network, yes — a compromise assumption is prudent given confirmed in-the-wild exploitation. CISA's required action references Forensics Triage Requirements that should be followed before returning the system to service. Rotate all secrets and audit all administrator accounts and API keys regardless of whether direct evidence of compromise is found.