Threat watch / Vulnerability

GPUStack: a forged header gives attackers admin over AI model servers, exploited in the wild

Vulnerabilities AI-related Medium Confirmed First seen 9 Nov 2026 · updated 11 Nov 2026

GPUStack, an open-source platform for running AI models on GPU clusters, trusted any request that appeared to come from the local machine. A request with a forged X-GPUStack-Real-IP: 127.0.0.1 header, or one passed through a reverse proxy on the same host, was treated as an unauthenticated admin, and an admin can register a custom inference backend that runs commands. The project fixed it in version 2.2.0 and says the flaw was exploited in the wild against a 2.1.1 deployment. CISA added it to its list of exploited flaws on November 9, 2026, according to The Hacker News.

Sector
Many sectors
Way in
Software flaw
Product or vendor
GPUStack
How it happened
A loopback trust flaw (CVE-2026-102895, CVSS 9.8): GPUStack before 2.2.0 treats requests that look local as trusted, so a forged X-GPUStack-Real-IP: 127.0.0.1 header, or any request relayed by a same-host reverse proxy, gets unauthenticated admin access. Code execution follows by creating a custom inference backend through POST /v2/inference-backends/from-yaml.[1][2] Checked 11 Nov 2026
What was stolen
Not public yet Checked 11 Nov 2026
Who was impacted
Organizations running GPUStack before 2.2.0 with the server reachable by the attacker. The advisory reports one exploited deployment, running 2.1.1, where attackers created custom backends running reverse shells; no count of victims is public.[1] Checked 11 Nov 2026
Threat actor
Not attributed
CVE
CVE-2026-102895

Timeline

  1. GPUStack publishes advisory GHSA-gj9m-f8ph-pmhr for the loopback trust flaw, fixed in 2.2.0, credited to Nikko Enggaliano, Wei Yunzhou, and Rui Yang and Haoyu Wang of Johns Hopkins University. [1]
  2. CISA adds CVE-2026-102895 to its Known Exploited Vulnerabilities catalog, according to The Hacker News; no attacker is named. GPUStack's advisory states that the flaw "has been observed exploited in the wild against a v2.1.1 deployment (attacker-created custom backends running reverse shells)", a report it credits to Javier Sixto. [1][2]

← Back in the news this week