CriticalCVE-2026-102895 4 Hunt LeadsIOCsRemote AttackerSANS Internet Storm Center

One Fake Header Makes You Admin on GPUStack, and Scanners Know It

SANS ISC honeypots are seeing exploit attempts against GPUStack's CVE-2026-102895, a CVSS 9.8 authentication bypass that turns a forged loopback header into full admin and code execution on AI worker hosts.

One fake header is all it takes to become admin on an unpatched GPUStack server, and someone is now trying it across the internet. SANS Internet Storm Center’s Johannes Ullrich reports that ISC honeypots are seeing live exploit attempts against CVE-2026-102895, an authentication bypass in GPUStack, the open-source software that manages GPU clusters for AI workloads. The fix has been out since GPUStack 2.2.0 in late June, and CISA added the bug to its exploited list on Nov. 9, as we covered.

For defenders

The bug: CVE-2026-102895 (CVSS 9.8) affects GPUStack before 2.2.0. When force_auth_localhost is off, the default, GPUStack treats any request that seems to come from 127.0.0.1 as the built-in administrator. Per the GitHub advisory, that happens two ways: a client-supplied X-GPUStack-Real-IP: 127.0.0.1 header, or simply sitting behind a reverse proxy on the same host, where every request arrives from loopback with no header needed (T1190 (Exploit Public-Facing Application)).

What an attacker gets: full admin over the API: users, models, workers and settings; persistent admin API keys valid from anywhere (T1098 (Account Manipulation)); worker credentials and cloud-provider secrets; and code execution on worker hosts through POST /v2/inference-backends/from-yaml. The advisory says a 2.1.1 deployment was hit in the wild with attacker-created custom backends running reverse shells (T1059.004 (Command and Scripting Interpreter: Unix Shell)).

What ISC sees: plain GET / requests with User-Agent: Mozilla/5.0 and the forged loopback header, from 144.91.73[.]103. The same address has been probing for other AI services at /v1/models, /v1beta/models?key=x, /v1/messages and /models (T1595.002 (Active Scanning: Vulnerability Scanning)). EPSS is 0.56%; no public exploit is reported in our CVE data, though the request is trivial to reproduce. CISA’s federal deadline was today, Nov. 12.

Network
144.91.73[.]103 (scanning source seen by SANS ISC)

Request
X-GPUStack-Real-IP: 127.0.0.1 (forged header from an untrusted client)
POST /v2/inference-backends/from-yaml (admin-only backend creation used for code execution)

What it means for you

If you don’t run GPUStack, there’s nothing to do. If your team does, this is the kind of bug that turns an AI server into someone else’s computer, with its cloud keys thrown in. GPU servers are expensive, so attackers like them for crypto mining as much as for the data on them.

Do this now

  1. Upgrade GPUStack to 2.2.0 or later now.
  2. Until you can, set force_auth_localhost=true, strip X-GPUStack-Real-IP at your proxy, and keep the API off the internet.
  3. If it was exposed, list admin API keys, users and inference backends, and remove anything you didn’t create.
  4. Rotate cloud-provider and worker credentials stored in GPUStack, then check worker hosts for reverse shells.

Source article: Exploit Attempts for GPUStack Authentication Bypass (CVE-2026-102895)
Publisher: SANS Internet Storm Center (Johannes Ullrich)
Published: Nov. 12, 2026
URL: https://isc.sans.edu/diary/33514

Executive Summary

SANS Internet Storm Center honeypots are recording exploitation attempts against CVE-2026-102895 (CVSS 9.8), an authentication bypass in GPUStack, the open-source GPU cluster manager used to serve AI models, by a source not attributed to any named actor. In GPUStack before 2.2.0, get_current_user promotes any request whose source looks like loopback to the built-in administrator while force_auth_localhost is false, the default; the source is spoofable with a client-supplied X-GPUStack-Real-IP: 127.0.0.1 header, and in the common deployment behind a same-host reverse proxy every remote request already arrives from 127.0.0.1. The fix shipped in 2.2.0 at the end of June 2026, the GitHub advisory followed on July 31, the CVE was assigned on Nov. 4 and CISA added it to the Known Exploited Vulnerabilities catalog on Nov. 9 with a Nov. 12 federal deadline. The advisory reports in-the-wild exploitation of a 2.1.1 deployment in which attackers created custom inference backends that ran reverse shells; with admin access an attacker can enumerate and change users, models, workers and configuration, mint persistent admin API keys valid from any source IP, read worker and cloud-provider secrets, and execute code on worker hosts through POST /v2/inference-backends/from-yaml. The requests ISC sees are minimal GET / probes with User-Agent Mozilla/5.0 and the forged header from 144.91.73[.]103, which has also been scanning for other AI API endpoints. The likely objectives are control of costly GPU capacity and theft of cloud credentials and models. Hunt leads, false positive expectations and indicators are detailed below; no detection rules have been published.

ATT&CK chain: Reconnaissance → Initial Access → Persistence → Credential Access → Execution

Hunt Lead Grid

Hunt Lead Level Hunt Lead Name ATT&CK Risk of False Positives
1 (Junior) ISC Scanning Source and AI Endpoint Probes T1595.002 (Active Scanning: Vulnerability Scanning) Low. The address is the scanning source ISC published and the probed paths are AI API routes; legitimate clients of your own AI services come from known networks, so the lead anchors on this source and on unauthenticated probes from the internet.
2 (Pro) Forged X-GPUStack-Real-IP Loopback Header T1190 (Exploit Public-Facing Application) Low. No legitimate external client sends X-GPUStack-Real-IP set to a loopback address; the only overlap is an internal gateway that sets the header itself, which can be allowlisted by its source address.
2 (Pro) Unexpected Admin API Keys, Users and Inference Backends T1098 (Account Manipulation), T1190 (Exploit Public-Facing Application) Medium. Administrators create keys and backends as normal work; the lead anchors on creations attributed to the built-in admin through a loopback source, outside change windows, or with full [“*”] scope.
3 (Elite) Reverse Shell from a GPUStack Worker After Backend Creation T1059.004 (Command and Scripting Interpreter: Unix Shell) Medium. Inference backends legitimately start containers and Python processes that reach model registries; the lead correlates a new from-yaml backend with a shell child process and an outbound connection to a destination outside the model-download allowlist.

ISC Scanning Source and AI Endpoint Probes

The exploit attempts ISC recorded are simple HTTP GET requests to / with Accept-Encoding: identity, User-Agent: Mozilla/5.0 and X-GPUStack-Real-IP: 127.0.0.1, all from 144.91.73[.]103. Over the preceding couple of days the same address also requested /v1/models, /v1beta/models?key=x, /v1/messages and /models, paths used by OpenAI-compatible, Gemini-style and Anthropic-style model APIs, which suggests a scanner looking broadly for exposed AI services rather than GPUStack alone. The detection logic is a retrospective search of web, proxy and load-balancer logs for any request from 144.91.73[.]103, and a broader search for unauthenticated requests from internet sources to those four paths on hosts that serve AI workloads, ranked by sources that also sent the X-GPUStack-Real-IP header.

Use reverse proxy and web server access logs (nginx, Caddy, Traefik), cloud load-balancer and WAF logs, and firewall logs for GPU hosts. Look back to at least late July 2026, when the advisory became public and exploitation was first reported. Scope to GPUStack servers and any other model-serving endpoints such as vLLM, Ollama or LiteLLM gateways. A single scanning address is a weak, fast-ageing indicator and a clean result proves little, so treat a hit as evidence of exposure that should send you to the header and admin-activity leads rather than as proof of compromise. Blocking the address is harmless but not a fix; removing the service from the internet is. A hit followed by any 200 response to an admin route should be escalated.

Hunt Lead Level: 1 (Junior Analyst)

ATT&CK: Reconnaissance → T1595.002 (Active Scanning: Vulnerability Scanning)

Sources: “Exploit Attempts for GPUStack Authentication Bypass (CVE-2026-102895)”, SANS Internet Storm Center, Nov. 12, 2026, https://isc.sans.edu/diary/33514 (source article)

Forged X-GPUStack-Real-IP Loopback Header

GPUStack’s client_ip_getter trusted the X-GPUStack-Real-IP header, which is meant to carry the real client address from a trusted proxy, and its authentication code treated a loopback value as the administrator. An attacker therefore only has to add X-GPUStack-Real-IP: 127.0.0.1 (or another loopback form) to any request to act as admin on versions before 2.2.0. The detection logic is a header rule at the edge: alert on any inbound request from a non-internal source carrying X-GPUStack-Real-IP, and at critical severity when its value is 127.0.0.0/8, ::1 or localhost, regardless of the destination path. Because the advisory’s second path, a same-host reverse proxy, needs no header at all, also flag any GPUStack deployment where the proxy forwards to 127.0.0.1 and force_auth_localhost is not set to true, as an exposure finding rather than an alert.

Implement the rule in the WAF or reverse proxy (log the header explicitly, since most default access-log formats omit custom headers) and, where available, in network IDS with HTTP header inspection. Add header logging now and review whatever history exists. Scope to every GPUStack instance and to the proxies in front of it. The only expected false positive is an internal gateway that sets the header deliberately; allowlist that gateway by source address and have it overwrite, never append to, client-supplied values. Detecting the header is the wrong place to stop: the durable controls are upgrading to 2.2.0 and stripping the header at the edge. Any internet-sourced request with a loopback value that received a 2xx response should be treated as a successful exploit and escalated, with the admin-activity lead run immediately for the same time window.

Hunt Lead Level: 2 (Pro)

ATT&CK: Initial Access → T1190 (Exploit Public-Facing Application)

Sources: SANS ISC (above), source article; “Authentication bypass via trusted loopback source (reverse-proxy TCP peer or X-GPUStack-Real-IP header) grants unauthenticated admin” (GHSA-gj9m-f8ph-pmhr), GPUStack, published July 31, 2026, https://github.com/gpustack/gpustack/security/advisories/GHSA-gj9m-f8ph-pmhr (additional source, same vulnerability)

Unexpected Admin API Keys, Users and Inference Backends

According to the GPUStack advisory, an attacker who bypasses authentication gains full administrative access and can create users, change models, workers and configuration, read worker credentials and cloud-provider secrets, and mint admin API keys with scope [“*”] that keep working from any source IP after the bug is patched. That last point is what makes this a persistence problem: upgrading GPUStack does not revoke keys an attacker already created. The detection logic is a review and ongoing alert on GPUStack’s own state and logs: every API key with full scope, every user and every inference backend, each with its creation time and creator, flagged when it was created by the built-in admin from a loopback source, created outside a change window, or does not match an owner on record; plus any read of worker or provider credential endpoints by such a session.

Use the GPUStack application logs and database (users, API keys, inference backends, workers), proxy logs for POST requests to key, user and backend routes, and the cloud provider’s own audit logs (CloudTrail, Azure Activity, GCP Audit) for use of any credentials GPUStack stores. Review all objects regardless of age, and look back in logs to late July 2026. Scope to every GPUStack instance that was ever reachable from untrusted networks. Normal administrator activity is the false positive source; confirm each unknown object with its supposed owner rather than suppressing by account, since the attacker acts as the built-in admin. Deleting only the attacker’s API key is the wrong scope of response: assume every secret GPUStack held was read and rotate cloud keys and worker credentials. Unknown full-scope keys or backends should be escalated as confirmed compromise.

Hunt Lead Level: 2 (Pro)

ATT&CK: Persistence → T1098 (Account Manipulation); Initial Access → T1190 (Exploit Public-Facing Application)

Sources: GHSA-gj9m-f8ph-pmhr, GPUStack, July 31, 2026, https://github.com/gpustack/gpustack/security/advisories/GHSA-gj9m-f8ph-pmhr (additional source, same vulnerability); SANS ISC (above), source article

Reverse Shell from a GPUStack Worker After Backend Creation

The advisory states that the one confirmed in-the-wild compromise, against GPUStack 2.1.1, used attacker-created custom inference backends that ran reverse shells, through the admin-only POST /v2/inference-backends/from-yaml route that lets an administrator define how a backend is launched on worker hosts. The command lines and destinations used in that incident have not been published. The correlation to build joins three events per worker host: a POST to /v2/inference-backends/from-yaml (or a new backend object) on the GPUStack server; a process started by the GPUStack worker or its container runtime that is a shell (sh, bash, dash) or a network utility (nc, ncat, socat, curl or wget piped to a shell, python -c with socket); and an outbound TCP connection from that process tree to an address outside the worker’s normal destinations, such as model registries, container registries and the GPUStack server.

Use EDR or auditd process execution data on worker hosts, container runtime events (Docker, containerd, Kubernetes audit logs if workers run in a cluster), NetFlow or firewall egress logs, and the GPUStack server’s proxy logs for the backend route. Baseline 30 days of child processes and destinations per worker, which on a dedicated inference host should be a short list, and look back to July 2026. Scope to GPU worker hosts first. Legitimate backends start Python and container processes and download weights, so anchor on interactive shells and on egress to destinations outside the allowlist rather than on process names alone. Killing the shell is not enough while the malicious backend definition remains; remove the backend, rotate the worker’s credentials and rebuild the host if a shell ran. Default-deny egress for GPU workers makes this detection far quieter and is the better long-term control, and the pattern generalizes to other AI-serving platforms that let administrators define launch commands.

Hunt Lead Level: 3 (Elite)

ATT&CK: Execution → T1059.004 (Command and Scripting Interpreter: Unix Shell)

Sources: GHSA-gj9m-f8ph-pmhr, GPUStack, July 31, 2026, https://github.com/gpustack/gpustack/security/advisories/GHSA-gj9m-f8ph-pmhr (additional source, same vulnerability)

Hash Appendix

No file hashes were published in the source article or the GPUStack advisory; this is an authentication-bypass vulnerability and no attacker payloads have been released.

References Appendix

  • “Exploit Attempts for GPUStack Authentication Bypass (CVE-2026-102895)”, SANS Internet Storm Center (Johannes Ullrich), Nov. 12, 2026, https://isc.sans.edu/diary/33514 (source article; observed requests and scanning source)
  • “Authentication bypass via trusted loopback source (reverse-proxy TCP peer or X-GPUStack-Real-IP header) grants unauthenticated admin” (GHSA-gj9m-f8ph-pmhr, CVE-2026-102895), GPUStack, July 31, 2026, https://github.com/gpustack/gpustack/security/advisories/GHSA-gj9m-f8ph-pmhr (root cause, impact, in-the-wild reverse shells, workarounds)
  • CISA Known Exploited Vulnerabilities catalog entry for CVE-2026-102895, CISA, Nov. 9, 2026, https://www.cisa.gov/known-exploited-vulnerabilities-catalog?field_cve=CVE-2026-102895
  • MITRE ATT&CK Enterprise technique pages linked above, MITRE, https://attack.mitre.org/

CVE at a glance

CVE-2026-102895CriticalKEVGPUStack
CVSS
9.8 Critical (v3.1)
EPSS
0.5% chance of exploitation in the next 30 days (higher than 44.4% of CVEs)
Exploited in the wild
Yes, on CISA’s KEV list since Nov 9, 2026
Public exploit
None reported in public exploit databases

Full CVE details

Data from NVD, FIRST EPSS and CISA KEV, refreshed automatically. We report whether public exploit code exists but never link to it: fake proof-of-concepts that infect researchers are common.

More from the Roundup