A default that quietly handed out admin rights
GPUStack is an open-source GPU cluster manager that thousands of teams install through pip or Docker Hub to serve AI models across their own hardware, often as a lightweight alternative to standing up a full commercial MLOps platform. In versions before 2.2.0, any request arriving from a loopback address, 127.0.0.1, was automatically granted built-in administrator rights, and the configuration setting that disables this behavior, force_auth_localhost, defaults to false out of the box. That default made the loopback check trivial to defeat rather than genuinely protective, since the check relied on a signal, the apparent source address, that turns out to be easy for anyone with network access to the host to manipulate.
Two paths let a remote attacker satisfy that loopback check without actually being on the local machine. A reverse proxy running on the same host as GPUStack, which is an extremely common deployment pattern for teams exposing an internal tool through a web-facing endpoint, will make every outside request appear to originate from 127.0.0.1 by the time GPUStack sees it, since the proxy itself is the one making the internal connection. Alternatively, an attacker could simply send a request with the header X-GPUStack-Real-IP set to 127.0.0.1, which GPUStack trusted without any verification that the header actually reflected the request's true origin. Either path converts an unauthenticated network request into full administrative access, with no credential, token, or prior foothold required.
What admin access on a GPU cluster actually buys an attacker
Administrator rights on GPUStack are not a narrow permission. They expose cloud API keys configured for the cluster, allow the attacker to mint additional admin tokens for persistence, and permit pushing YAML configuration that executes arbitrary code on worker nodes across the cluster. In practice, that means an attacker who exploits this flaw does not just gain visibility into an AI serving environment, they gain remote code execution across every machine the cluster manager controls.
Researcher Javier Sixto of ulex.fr reported in-the-wild exploitation against a GPUStack deployment running version 2.1.1, in which attackers used custom backends to establish reverse shells on the compromised infrastructure, giving them an interactive command channel into the cluster rather than a single automated payload drop. Analysts credited with supporting work on the vendor advisories, Nikko Enggaliano, Wei Yunzhou, Rui Yang, and Haoyu Wang, round out a research effort that moved from discovery to a confirmed live compromise fast enough that GPUStack's maintainers and CISA both treated it as an urgent, already-active threat rather than a theoretical disclosure. That is a meaningfully different posture than most newly published vulnerabilities carry, and it shortens the window security teams have to treat this as optional reading.
A second, wider flaw in the same product
GPUStack's vendor advisory also covers CVE-2026-102894, rated CVSS 9.1, which affects a much broader version range, 0.7.1 through 2.2.2, covering nearly every release the project has ever shipped and every deployment that has not already moved to the newest patch. This flaw involves GPUStack's SAML callback route skipping signature verification entirely, which in principle would let an attacker forge SAML assertions and authenticate as any user, administrator included, without needing valid credentials of any kind. No exploitation of this second flaw has been confirmed in the wild yet, but its presence in the same advisory, in the same product, tied to the same underlying authentication architecture as the loopback bypass, is a signal that GPUStack's identity handling needs a full architectural review rather than a single point fix for whichever bug researchers happen to find first.
GPUStack fixed the loopback bypass in version 2.2.0 by removing the auto-admin rule entirely, and added SAML signature verification in version 2.2.2. For teams that cannot upgrade immediately, the vendor's interim guidance includes enabling force_auth_localhost explicitly, stripping the X-GPUStack-Real-IP header at the reverse proxy layer before it reaches GPUStack, and blocking external access to the /auth/saml/callback endpoint, none of which requires a code change but all of which require someone to actually revisit a configuration most teams set once at install time and never touched again.
Why this belongs in your AI governance conversation
AI infrastructure tooling like GPUStack typically enters an organization the way developer tools always have: an engineering team installs it to solve an immediate problem, and it goes into production before a formal security review ever happens, often starting as a proof of concept on a single workstation before quietly becoming the system three other teams now depend on. That path bypasses the procurement and architecture review gates that would normally catch a default-admin-on-loopback design decision before it reached a live environment serving real models and real cloud credentials, which is precisely how a research convenience tool ends up as unreviewed production infrastructure.
CISA's decision to add this flaw to its Known Exploited Vulnerabilities catalog on November 9, in the same batch as three other actively exploited bugs spanning completely unrelated product categories, is a useful marker for how far AI infrastructure has moved from experimental to production-critical in the government's own risk calculus. We think CISOs should read that inclusion less as news about one product and more as confirmation that the AI serving layer is now squarely inside the attack surface their teams are accountable for, whether or not it went through a normal onboarding process, and whether or not the team that stood it up ever filed a ticket with security to let them know it existed in the first place.
What this means for your roadmap
Start by inventorying every self-hosted AI or GPU serving tool in your environment, including anything an engineering or data science team stood up independently, since this class of tooling routinely skips the review gates applied to conventional enterprise software. For any GPUStack deployment, upgrade to 2.2.2 or later immediately, and in the meantime verify force_auth_localhost is explicitly enabled and that your reverse proxy layer strips any client-supplied real-IP headers before they reach the application.
Treat admin access to AI infrastructure with the same severity rating you would assign to admin access on a core identity system, given that it now routinely carries cloud credentials and code execution rights across a cluster rather than visibility into a single application. Finally, build AI and ML serving platforms into your standard vulnerability management and procurement review cycle going forward, rather than treating them as a special category that engineering teams manage independently of the rest of the security program, since this flaw shows that independence is exactly how a critical system goes unreviewed until it is already compromised.



