← All articles

GPUStack Bypass Flaw Threatens Swiss AI Ops 2026

An unauthenticated header-spoofing flaw in the open-source GPUStack GPU cluster manager now sits in CISA's Known Exploited Vulnerabilities catalog, with live attacks already recorded.

A critical authentication bypass in GPUStack, open-source software used to manage GPU clusters for AI model serving, is being actively exploited in the wild. CISA added CVE-2026-102895 to its Known Exploited Vulnerabilities catalog on 2026-11-09, setting a mandatory patch deadline of 2026-11-12 S1. The flaw, rated CVSS 9.8, lets unauthenticated attackers impersonate the built-in administrator account, create persistent API keys, extract worker and cloud-provider secrets, and execute arbitrary code S1. SANS ISC has confirmed live exploit attempts against exposed instances S2.

How a Single Header Defeats GPUStack's Authentication

The root cause sits in gpustack/api/auth.py, where requests that appear to originate from the loopback address are automatically treated as the built-in administrator. This behaviour is governed by a setting called force_auth_localhost, which defaults to false, meaning the dangerous shortcut is active out of the box in GPUStack versions prior to 2.2.0 S1. Operators who deployed GPUStack without explicitly hardening this setting are exposed by default, not through misconfiguration but through the software's own baseline behaviour.

Exploitation is trivial because it does not require credential theft, session hijacking, or any cryptographic work. An attacker simply sets the X-Gpustack-Real-IP header to 127.0.0.1 in an otherwise ordinary HTTP request, and the application accepts the spoofed origin as proof of local, trusted access S2. This is a textbook case of origin validation failure: the server trusts a client-supplied header as if it were a network-level fact, collapsing the entire authentication model into a single string comparison that any attacker can forge.

From Spoofed Header to Full Administrative Takeover

Once the spoofed header is accepted, the attacker inherits full administrator privileges inside GPUStack. Documented consequences include arbitrary changes to users and deployed models, creation of persistent administrator API keys that survive beyond the initial session, and disclosure of worker and cloud-provider secrets held by the platform S1. For organisations running GPU infrastructure tied to cloud billing accounts or model-serving pipelines, secret disclosure alone can cascade into lateral compromise well beyond the GPUStack instance itself.

The most severe consequence is remote code execution through the POST /v2/inference-backends/from-yaml endpoint, which an authenticated administrator can use to register new inference backends S1. Because the authentication bypass grants that same administrative role to anyone who spoofs the header, this endpoint becomes a direct path to code execution on the underlying host. The vulnerability is formally classified under CWE-287 for improper authentication, CWE-290 for authentication relying on hard-coded trust assumptions, and CWE-346 for origin validation error, reflecting that this is a layered design failure rather than a single coding mistake S1.

Live Exploitation Is Already Under Way

SANS ISC's published findings go beyond theoretical risk: reconnaissance and exploit attempts have been observed from a specific source address, 144.91.73.103, probing exposed GPUStack instances S2. The requests target model and inference endpoints including /v1/models, /v1beta/models, /v1/messages, and /models, suggesting the attacker is mapping which API surfaces are reachable before or alongside attempting the authentication bypass itself S2.

This pattern is consistent with automated, internet-wide scanning rather than a targeted campaign against a single victim. That makes exposure time the dominant risk factor: any internet-facing GPUStack deployment running a vulnerable version is a candidate target regardless of its owner's size or sector. The combination of a trivially forged header, publicly known exploit mechanics, and confirmed scanning activity means unpatched instances should be assumed to be under active probing right now, not merely theoretically at risk.

What Swiss AI and Financial Operations Should Do Now

Swiss financial institutions and AI-intensive enterprises using GPUStack for model serving or GPU provisioning face direct exposure if any instance is reachable from the internet and running a version prior to 2.2.0. The risk is not abstract: successful exploitation yields unauthenticated code execution and access to secrets tied to worker nodes and cloud providers, which is a serious concern for any organisation handling regulated data or operating infrastructure subject to supervisory expectations around operational resilience. Teams should treat the CISA catalog addition and its associated patch deadline as the operative deadline for their own remediation planning, even where CISA's authority does not directly apply to their organisation S1.

The immediate task is inventory, not just patching: many organisations do not have a precise map of every internet-reachable service running GPUStack, particularly where it was deployed by data science or ML engineering teams outside the usual change-management process. Closing this flaw requires both a version check and a configuration check, since the vulnerable default persists even in deployments that administrators believed were adequately isolated. As GPU infrastructure becomes a more central and more targeted part of enterprise technology stacks, the lesson for Swiss security teams is that AI tooling deserves the same perimeter scrutiny, patch discipline, and exposure monitoring long applied to traditional production systems.

◆ Key Takeaway

Immediately identify any internet-exposed GPUStack instance running a version prior to 2.2.0, confirm force_auth_localhost is not left at its insecure default, and patch or isolate it today given confirmed active exploitation.

  • Inventory all GPUStack deployments across model-serving and GPU-provisioning environments, including those stood up outside formal IT change control.
  • Verify the installed GPUStack version and prioritise upgrading any instance running a version prior to 2.2.0.
  • Explicitly set force_auth_localhost to true rather than relying on the insecure default, even after patching.
  • Audit perimeter and firewall rules to confirm GPUStack management and API endpoints are not reachable from the public internet.
  • Review logs for requests containing a spoofed X-Gpustack-Real-IP header and for traffic associated with IP 144.91.73.103 or similar reconnaissance patterns.
  • Rotate any API keys, worker credentials, and cloud-provider secrets accessible to GPUStack instances that were exposed prior to patching.
  • Treat the CISA Known Exploited Vulnerabilities deadline as the internal remediation benchmark and document completion for audit and supervisory purposes.