GPUStack CVE-2026-102895 is a critical authentication bypass vulnerability. It enables unauthenticated attackers to gain full administrative access to GPUStack deployments. Furthermore, active exploitation has been confirmed in the wild, making immediate patching essential for organizations running AI inference infrastructure.
What Happened: GPUStack Authentication Bypass Under Active Exploitation
On November 9, 2026, the U.S. Cybersecurity and Infrastructure Security Agency (CISA) added CVE-2026-102895 to its Known Exploited Vulnerabilities catalog. Consequently, federal agencies must patch by November 24, 2026, under Binding Operational Directive 26-04.
The vulnerability affects GPUStack versions prior to 2.2.0. In addition, threat actors have already exploited the flaw against real-world deployments. Javier Sixto of Ulex Innovative Systems reported active in-the-wild exploitation. Attackers created custom inference backends running reverse shells after gaining admin access.
Technical Details of the GPUStack Authentication Bypass
The flaw resides in the get_current_user function within gpustack/api/auth.py. Moreover, the code automatically promotes any request originating from a loopback address to the built-in administrator account. This behavior triggers when force_auth_localhost is set to False, which is the default configuration.
Attackers can exploit this weakness through two distinct paths. First, a same-host reverse proxy configuration causes every remote request to arrive with a TCP peer of 127.0.0.1. Therefore, the loopback auto-admin branch fires without any credentials. Second, attackers can forge an X-GPUStack-Real-IP: 127.0.0.1 header to trigger the same privileged path.
- CVE ID: CVE-2026-102895
- Severity: Critical (CVSS score: 9.8)
- Affected versions: GPUStack prior to 2.2.0
- Attack vector: Network — unauthenticated
- Privileges required: None
- Impact: Full administrative takeover, code execution on worker hosts
- Exploitation: Active in the wild
How the Attack Chain Works
- The attacker sends a crafted request to the GPUStack API endpoint.
- The request appears to originate from 127.0.0.1 through a reverse proxy or forged header.
- GPUStack assigns the built-in administrator account to the session automatically.
- The attacker creates persistent admin API keys with global scope.
- Using admin privileges, the attacker deploys a malicious inference backend via
POST /v2/inference-backends/from-yaml. - The custom backend executes a reverse shell on worker hosts, enabling persistent access.
Business and Operational Impact
GPUStack is an open-source GPU cluster manager widely used for AI and machine-learning workloads. Moreover, the platform manages sensitive worker credentials and cloud-provider secrets. A successful compromise exposes the entire AI infrastructure stack to attackers.
The impact extends beyond a single application. For example, attackers can read worker credentials, modify models, exfiltrate training data, and pivot to connected cloud environments. Organizations running GPUStack in production face immediate risk if they operate unpatched instances behind reverse proxies.
- Data theft: Access to model weights, training datasets, and inference logs.
- Credential exposure: Worker secrets and cloud-provider API keys become readable.
- Compute abuse: Attackers can hijack GPU resources for cryptomining or unauthorized inference.
- Lateral movement: Compromised workers serve as launch points into broader networks.
- Compliance risk: Unauthorized access to AI systems may trigger breach notification requirements.
Mitigation and Recommendations
Immediate Actions for Defenders
- Upgrade immediately. Install GPUStack 2.2.0 or later. The developers removed the loopback auto-admin path and hardened IP-forwarding logic.
- Enable forced authentication. Set
force_auth_localhost=trueto require authentication for all requests, including loopback traffic. - Audit API keys. Review existing admin API keys for unauthorized creations. Revoke any suspicious keys immediately.
- Inspect inference backends. Check for unexpected custom backends deployed through the
from-yamlendpoint. - Review access logs. Look for requests with
X-GPUStack-Real-IP: 127.0.0.1or unusualPOST /v2/inference-backends/from-yamlentries.
Network and Architectural Hardening
- Strip
X-GPUStack-Real-IPheaders at the edge proxy before they reach GPUStack. - Restrict GPUStack API access to trusted networks, VPN segments, or authenticated gateways.
- Segment worker nodes from production networks to limit lateral movement.
- Monitor GPU utilization anomalies that may indicate unauthorized workloads.
Bottom line: CVE-2026-102895 is a critical, actively exploited vulnerability that grants unauthenticated admin access to GPUStack. Organizations must patch to version 2.2.0, audit their environments for signs of compromise, and restrict API exposure to trusted networks.
Incident Summary
| CVE ID / Incident: | CVE-2026-102895 |
| Affected Systems: | GPUStack prior to version 2.2.0 |
| Disclosure Date: | July 31, 2026 (active exploitation confirmed November 2026) |
| Patch Status: | Available — upgrade to GPUStack 2.2.0 or later |
| CVSS Score: | 9.8 (Critical) |
| CISA KEV Addition: | November 9, 2026 |
| Federal Patch Deadline: | November 24, 2026 |
References
- CISA, “CISA Adds Four Known Exploited Vulnerabilities to Catalog,” November 9, 2026. https://www.cisa.gov/news-events/alerts/2026/11/09/cisa-adds-four-known-exploited-vulnerabilities-catalog (accessed November 12, 2026).
- GPUStack, “Authentication bypass via trusted loopback source (reverse-proxy TCP peer or X-GPUStack-Real-IP header) grants unauthenticated admin,” GitHub Security Advisory GHSA-gj9m-f8ph-pmhr, July 31, 2026. https://github.com/gpustack/gpustack/security/advisories/GHSA-gj9m-f8ph-pmhr (accessed November 12, 2026).
- The Hacker News, “CISA Warns of Active Exploitation of OpenSSH, GPUStack, FreeSWITCH, and IBM Flaws,” November 10, 2026. https://thehackernews.com/2026/11/cisa-warns-of-active-exploitation-of.html (accessed November 12, 2026).