GPUStack CVE-2026-102895: when trusting 127.0.0.1 turns into remote RCE-like admin access
GPUStack’s auth bypass trusting 127.0.0.1 (CVSS 9.8): how remote admin access was possible, why it landed in CISA KEV, and how to upgrade.

GPUStack is an open-source platform for pooling GPUs across machines and serving models on top of them. It’s practical for teams running LLMs on their own hardware: add GPU nodes, pull and deploy models, access via API.
CVE-2026-102895, published November 4, 2026, describes a CVSS 9.8 authentication bypass in the platform. On November 9 it was added to CISA’s Known Exploited Vulnerabilities (KEV) catalog—meaning there is evidence of active exploitation. This post explains the technical mechanism and what defenders should do.
Quick summary
| Field | Value |
|---|---|
| CVE | CVE-2026-102895 |
| CVSS 3.1 | 9.8 (Critical) |
| Vector | AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H |
| Weakness | CWE-287, CWE-290, CWE-346 |
| Status | Added to CISA KEV (active exploitation indicated) |
What went wrong
The authentication logic treated requests as trusted based on assumptions tied to localhost (127.0.0.1) in contexts where that trust could be influenced by attacker-controlled request routing/headers/proxy behavior (exact conditions depend on deployment). The result: paths intended for local/admin contexts could be reached remotely without proper authentication.
Impact
An unauthenticated remote attacker could gain administrative control over the GPUStack instance—accessing models, credentials/secrets managed by the platform, and potentially pivoting into connected infrastructure.
Mitigation
- Upgrade to the vendor-fixed version immediately.
- Restrict exposure: Do not expose management interfaces to the public internet; enforce network ACLs/VPN.
- Audit: Review logs for suspicious admin/API activity around the disclosure window; rotate tokens/keys if compromise is suspected.
Self-hosted AI platforms concentrate high-value assets (models, data, keys). Treat them as Tier-0 services.