GPUStack loopback trust grants remote requests administrator access
CISA added an exploited GPUStack authentication bypass to KEV. Upgrade to 2.2.0 and investigate administrator keys, backend changes and exposed secrets.

A remote request could reach the built-in GPUStack administrator account without a password when the server trusted a local-looking connection. CISA added the flaw, CVE-2026-102895, to its Known Exploited Vulnerabilities catalog on November 9. Teams running GPUStack for GPU clusters and AI model serving need to check their installed version and the path traffic takes into the management API.
The GPUStack maintainer advisory says attackers already exploited a 2.1.1 deployment and created custom inference backends that ran reverse shells. The advisory dates from July 31; CISA’s new KEV listing brings an explicit November 12 deadline for covered federal agencies. The CVE Program record lists releases before 2.2.0 as affected and 2.2.0 as fixed. The advisory credits independent reports by Nikko Enggaliano, Wei Yunzhou, and Rui Yang and Haoyu Wang, and credits Javier Sixto with reporting exploitation in the wild.
How the local shortcut failed
GPUStack’s get_current_user routine in gpustack/api/auth.py treated a request whose source appeared to be loopback as the built-in administrator when force_auth_localhost was false, the default before 2.2.0. The client_ip_getter made that local-source decision. A same-host reverse proxy could forward a remote user’s request to GPUStack with a loopback TCP peer, making ordinary remote traffic look local to the application. Alternatively, a client-controlled X-GPUStack-Real-IP header could supply the loopback value 127.0.0.1. The advisory says the same-host proxy route required neither a special header nor credentials.
The advisory lists changes to users, models, workers and configuration; persistent administrator API keys with wildcard scope *; access to worker credentials and cloud-provider secrets; and code execution on worker hosts through the administrator-only inference-backend import path /v2/inference-backends/from-yaml. The maintainers observed attacker-created reverse-shell backends in a 2.1.1 deployment; the other listed actions describe what the bypass can enable. CISA describes the issue as allowing unauthenticated administrative access and flags forensic triage for this KEV entry.
Recognize access that outlived the request
Start with GPUStack’s version and exposure map: identify each server below 2.2.0, its same-host proxy or gateway, and whether untrusted clients can reach the API. A successful attacker could create an administrator key that remains valid from other source IPs after the local-looking request ends. That changes the investigation: removing the proxy exposure alone leaves previously issued credentials usable. Review available API and reverse-proxy request records for management operations associated with unexpected source identities, creations of wildcard-scope admin keys, changes to users or models, and inference-backend additions. On affected servers, compare backend definitions and worker activity with approved changes; the reported incident involved backends running reverse shells. Check access to worker credentials and cloud-provider secrets and rotate any exposed values as part of incident response.
A reverse-proxy log can retain the external client address while GPUStack receives a loopback TCP peer. Correlate proxy requests with GPUStack events rather than treating the application’s local peer address as the real visitor. Logging coverage and retention vary by deployment; a clean search in one log cannot clear an exposed server whose administrator keys and worker state have not been reviewed.
Upgrade the server and constrain entry
The maintainers say, “Upgrade to 2.2.0 or later.” Version 2.2.0 removes the loopback auto-admin path and hardens client-IP and forwarding-header handling. Until that upgrade, set force_auth_localhost=true so even loopback requests require authentication. Configure the edge to strip client-supplied X-GPUStack-Real-IP, and keep the GPUStack API behind a VPN or edge authentication rather than exposing it to untrusted networks. Header stripping addresses the client-supplied route, while the same-host proxy peer route still needs the local-authentication setting or the upgraded code. If the API has a separate direct listener, include it in the exposure review: an edge rule on the proxy does not govern clients reaching that listener.
Verify authentication after the change
On each upgraded deployment, confirm the installed GPUStack release is 2.2.0 or later and inspect the proxy and gateway configuration for the header-stripping and access controls you intended. From a normal remote client through the same proxy, make a benign request to a management resource without credentials and verify that it receives an authentication challenge rather than administrator data. Repeat the access test on the direct API path if that path is reachable. Inventory administrator API keys and reconcile them against approved owners before relying on the new version to protect an already exposed installation. Revoke any unexplained keys. Preserve the related request records and backend configuration for incident response before removing a suspicious backend; then check whether its worker had access to credentials that need rotation.
The fixed release removes the trust shortcut for new requests. A key minted during earlier access persists until you find and revoke it; backend changes and worker secrets need their own investigation.