CISA added GPUStack’s CVE-2026-102895 to its Known Exploited Vulnerabilities catalog on November 9, 2026 with a due date of November 12, which is today. The flaw can give unauthenticated administrative access to GPUStack versions before 2.2.0 when a same-host reverse proxy or a forged header makes a remote request look local. Version 2.2.0 has been available since June 18. The vendor advisory reports exploitation against a v2.1.1 deployment, where the attacker created custom backends that ran reverse shells. It also says an attacker with this access can create administrator API keys that are valid from any source IP afterwards, and it does not say that upgrading revokes them. So there are three jobs, in the order CISA’s triage guidance puts them: copy the logs and the server’s data aside, get onto a fixed version, then go through what an unauthenticated admin could have left behind. S5 Labs has not tested any of it; what follows is read from the KEV feed, the GitHub advisories, the release notes and the v2.1.1 and v2.2.0 source. CISA KEV catalog entry · GHSA-gj9m-f8ph-pmhr · GPUStack releases
What the KEV entry asks for, and in what order
The catalog we read is version 2026.11.10, released November 10, and it has one GPUStack entry. CISA’s description lists user and model changes, persistent administrator API-key creation, disclosure of worker and cloud-provider secrets, and code execution through POST /v2/inference-backends/from-yaml. Ransomware use is marked “Unknown” and forensic triage “Yes”. The required action is to apply mitigations per the vendor’s instructions, in line with CISA’s BOD 26-04 and its Forensics Triage Requirements, or to stop using the product if mitigations are unavailable. KEV JSON feed
The directive is addressed to federal agencies; for everyone else the triage guidance is a usable template. CISA says the directive requires an adequate forensic triage analysis, while “the specific timeline below is not required”. Its baseline targets, counted from the KEV addition:
| Step | Target |
|---|---|
| Scoping | Within the first 2 hours |
| Preserve and collect evidence | 2 to 24 hours |
| Critical patching | 2 to 24 hours, applied only after required evidence is collected |
| Contain and control | 6 to 24 hours |
| Triage analysis | 24 to 48 hours |
| Escalation decision and report | 48 to 72 hours |
Source: CISA, BOD 26-04 implementation guidance. The page also says to avoid altering or remediating systems before evidence collection when possible. For GPUStack, our reading is to copy the logs and the server’s data aside before replacing anything.
How a proxy on the same host made every visitor the admin
Before 2.2.0, get_current_user in gpustack/api/auth.py promoted any request whose source counted as loopback to the built-in administrator account when force_auth_localhost was False, and False was the default. In the v2.1.1 source the branch fires when no credentials resolved a user and the client IP equals 127.0.0.1. GHSA-gj9m-f8ph-pmhr · auth.py at v2.1.1
The advisory names two ways a remote request ends up looking local.
| Path | What the advisory states | What the v2.1.1 code shows |
|---|---|---|
| Same-host reverse proxy | With proxy_pass http://127.0.0.1:<port>, every remote request reaches GPUStack with TCP peer 127.0.0.1. No header or credential is needed. The advisory says it verified this end to end. | Outside embedded gateway mode, the client IP is the TCP peer |
| Forged header | An X-GPUStack-Real-IP: 127.0.0.1 header can make the source look like loopback | The header is read only when gateway_mode is embedded; the default is auto, which resolves to one of four modes depending on environment |
The header path therefore appears limited to embedded gateway mode in 2.1.1, a qualifier the advisory leaves out. We did not test which mode a typical install resolves to, so find out what yours resolved to before treating the header as harmless. The proxy path needs no such condition, and the advisory describes that layout as the canonical self-hosted one.
GitHub rates the advisory Critical at CVSS 3.1 9.8. NVD lists that figure as a secondary score from GitHub and assigns none of its own. EPSS stood at 0.517% on November 12, around the 44th percentile. CISA lists the flaw as exploited, so do not use that number to rank it down. The public evidence of exploitation is the advisory’s account of one compromised deployment. None of the primary pages we read gives a victim count, a first-exploitation date or indicators of compromise. GHSA-gj9m-f8ph-pmhr · NVD, CVE-2026-102895
Version 2.3.1 is the one to run
Version 2.2.0 closes this CVE but none of the flaws fixed in the releases after it.
| Release | Date | Fixes | Still open on this release |
|---|---|---|---|
| 2.2.0 | June 18 | Loopback auto-admin bypass, CVE-2026-102895 (PR #5346, merged May 19) | SAML callback flaw; Kubernetes proxy authorization flaw GHSA-3hch-x692-6qj3 (High, 8.8, needs a non-admin login) |
| 2.2.2 | July 24 | SAML callback, CVE-2026-102894; missing authorization on several cluster and resource endpoints | Everything fixed in 2.3.x |
| 2.3.0 | October 23 | World-readable secret files GHSA-72j3-6rmw-742x; Kubernetes proxy path normalization GHSA-8wmp-hw5p-v2j7 | The 2.3.1 list |
| 2.3.1 | November 11 | Code execution during model configuration evaluation GHSA-gjf9-3p3r-9h9c (Critical, 9.1, needs permission to submit models); resource ownership changes GHSA-33f6-vcmh-2xfm; Hugging Face token and registry password disclosure GHSA-7345-vjfr-2hcx; others | Latest stable |
Sources: the v2.2.0, v2.2.2, v2.3.0 and v2.3.1 release notes and the repository’s advisory list. Nineteen of that list’s 22 advisories were published between October 26 and November 11, most of them authorization flaws that need a login.
Run 2.3.1, the current stable release. If you move a Helm or Kubernetes deployment across 2.2.2, its notes list the rename of GPUSTACK_USAGE_ROLLUP_TIMEZONE to GPUSTACK_TIMEZONE and of the worker Service from worker to gpustack-worker.
The SAML flaw is the one to pair with the loopback bypass, because it also needs no credentials. Per GHSA-62m9-f875-rqw6 (Critical, CVSS 3.1 9.1, affecting 0.7.1 and later, fixed in 2.2.2), the callback read a posted SAML response and issued a session without validating its signature, issuer, audience, destination or time. The advisory says SAML does not have to be configured, and the only precondition is network reach to /auth/saml/callback. Naming an existing user, an administrator for instance, yields that user’s session; otherwise a new non-admin user is created. The advisory offers no configuration workaround short of blocking that path at the proxy or ingress. This CVE is not in the 2026.11.10 KEV feed, and none of the primary pages we read reports exploitation.
The 2.2.0 release notes contain no security section and do not mention the loopback bypass. The repository advisory followed on July 31, the NVD record on November 4 and the GitHub Advisory Database entry on November 9. The vendor does not explain the sequence. Anyone who decided in June to stay on 2.1.x on the strength of those notes decided without this information.
What upgrading leaves in place
This section is our analysis; neither GPUStack nor CISA states it. The advisory says an attacker can create persistent admin API keys with scope ["*"], read worker credentials and cloud or provider secrets, and reach code execution on worker hosts through admin-only functions. The fix removed the loopback auto-admin path. Nothing in the advisory says it invalidates a key issued earlier, removes a backend already registered, or changes a secret already read. The Xinference 4.1.0 advisory (our coverage) said outright that upgrading leaves an attacker-created admin in place; GPUStack’s is silent on the question. No advisory reported exploitation of the Xinference flaws, but CISA lists this one as exploited.
What to check
These steps are S5 Labs’ judgment, built on the advisory’s impact list. The advisory itself lists only upgrading and three workarounds.
Exposure history. Establish whether any version before 2.2.0 sat behind a same-host reverse proxy, or ran with gateway_mode resolved to embedded, and for how long. pip show gpustack or the image tag gives the version. In the v2.1.1 config.py, auto becomes embedded on a server that is not inside Kubernetes and has no gateway_kubeconfig set. If either answer is yes, assume any remote client in that period could have acted as admin. Either way, copy the logs and the server’s data aside before you upgrade.
Admin users and API keys. Account for every administrator and every API key, and revoke what nobody can explain. We would delete and reissue every admin key rather than try to prove each one clean.
Custom inference backends. Compare registered backends with your change records and remove any you did not create. Attacker-created backends are what the one reported compromise used.
Worker hosts. These are the GPU machines themselves, the kind in our Local LLM Machines guide. Check process trees and outbound connections, and rebuild a worker you cannot vouch for.
Secrets. Rotate worker tokens and the cloud and provider credentials the server held. Before 2.3.0 the JWT signing key file could also be created world-readable under a permissive umask such as 022, and its holder can forge administrator sessions, so rotate it if attacker code may have run on the host that stores the server’s data.
If you cannot upgrade today. The advisory’s workarounds are force_auth_localhost=true, a proxy that strips client-supplied X-GPUStack-Real-IP and cannot present external traffic as loopback, and keeping the API behind a VPN or edge authentication. From 2.2.0 the flag does nothing: the v2.2.0 config.py labels it “(Deprecated, no-op)” and says localhost callers are always authenticated. Block /auth/saml/callback until you are on 2.2.2 or later.
We could not verify on a primary page what a from-yaml creation looks like in GPUStack’s logs, whether 2.2.x logs API-key creation, or the exact screens for listing and revoking keys. Test those on your own install before relying on log review alone.
Sources
- CISA: KEV catalog entry, KEV JSON feed, BOD 26-04 implementation guidance
- Loopback bypass: GHSA-gj9m-f8ph-pmhr, PR #5346, NVD record, auth.py at v2.1.1, config.py at v2.1.1 and v2.2.0
- SAML callback: GHSA-62m9-f875-rqw6
- GPUStack releases: v2.2.0, v2.2.2, v2.3.0, v2.3.1, advisory list
Related S5 Labs coverage: Anthropic’s five exploit chains: the patch checklist · How AI finds vulnerabilities
