On this page (8 sections)
GPUStack, an open-source GPU cluster manager used for AI model serving, has a critical authentication flaw. According to the GitHub advisory, the SAML callback endpoint builds a login session from posted XML without ever verifying it. Anyone who can reach the endpoint over the network can sign in as an existing user, including an administrator.
The bug affects every deployment on a vulnerable version, not only those that use SAML single sign-on. If you run GPUStack, especially on a network that is reachable from the internet or from many internal users, treat this as an upgrade-today item.
What happened
The vulnerable endpoint is /auth/saml/callback. Per the advisory, it Base64-decodes the posted XML, reads the NameID and attributes, finds or creates the matching user, and issues a gpustack_session JWT. It skips the library step that is supposed to check the signature, issuer, audience, destination and time conditions (OneLogin_Saml2_Auth.process_response).
In plain terms, the server believes whatever identity the request claims. SAML is designed so that an identity provider signs a statement and the application verifies that signature. Here the verification never happens, so the signature is decoration.
What an attacker can do
- Name an existing account, such as an administrator, and receive that user's session. This is full account takeover.
- Name an unknown identity and have a new non-administrator account provisioned.
- Replay a captured signed response indefinitely, because Response and Assertion IDs are never consumed.
The advisory notes that administrator access gives full control of the deployment and the infrastructure it manages. For a GPU cluster manager, that can mean the models, credentials and compute nodes behind it. No credentials and no user interaction are required, and the only precondition is network reachability to the endpoint.
Who is affected
Every GPUStack installation from 0.7.1 up to, but not including, 2.2.2. The key point from the advisory is that disabling or never configuring SAML does not protect you: the callback runs regardless. There is no effective configuration-level workaround.
Exposure is highest for instances published to the internet, shared lab or team environments, and any deployment where the management port is reachable from untrusted networks. Instances restricted to a small, trusted network are at lower risk but still need the update, since any internal user or compromised machine could abuse the flaw.
How to check your exposure
Start with the installed version. This read-only command, run in the Python environment where GPUStack is installed, prints the package version. Anything from 0.7.1 to 2.2.1 is vulnerable.
pip show gpustackIf you run GPUStack in a container or other packaged form, check the image or release tag you deployed against the fixed release, v2.2.2. Then look at who can reach the service. If you have a reverse proxy or ingress in front, search its access logs for requests to the callback path. This read-only search lists matches; adjust the log path to your setup.
grep "/auth/saml/callback" /var/log/nginx/access.logIf you do not use SAML, any request to that path is suspicious. If you do use SAML, compare the source addresses and timing with your real login activity. Also review the user list for accounts nobody created, since an unknown identity results in a newly provisioned account.
What to do now
- Upgrade GPUStack to 2.2.2 or later, following your normal deployment method.
- If you cannot upgrade immediately, block external access to /auth/saml/callback at your reverse proxy or ingress, as the advisory recommends.
- Restrict network access to the GPUStack web interface and API to trusted networks or a VPN.
- Review the user list and remove accounts you do not recognize, especially new non-admin users.
- Check administrator activity and recent sessions for anything you cannot explain, and rotate secrets and API keys stored in or managed by the deployment if you suspect abuse.
- Search proxy and application logs for requests to /auth/saml/callback from unexpected addresses.
To upgrade a pip-based installation, run the command below in the GPUStack environment, then restart the service the way you normally do and re-check the version.
pip install --upgrade gpustack
pip show gpustackTemporary blocking at the proxy
The advisory only says to block network access to the callback until you upgrade. If you use nginx as the reverse proxy, a location block like this one denies the path. Add it to the relevant server block and reload nginx. It is a stopgap, not a substitute for the update.
location = /auth/saml/callback {
return 403;
}What the fix changes
Version 2.2.2 now validates the response through the proper verification call and rejects anything that fails the signature, issuer, audience, destination or time checks. It also makes the SAML and OIDC login and callback routes run only when that provider is the configured active one, so a deployment that has not enabled a provider no longer parses attacker-supplied identity payloads. The changes are in the linked pull request 5753 and pull request 5923.
Detection and follow-up
A forged login looks like a normal session, so logs and account state are your best evidence. After patching, keep an eye on these signals:
- Any hit on the callback path when SAML is not in use.
- New user accounts or administrator logins from unfamiliar addresses.
- Unexpected changes to models, workers, API keys or cluster settings.
- Sessions created outside your identity provider's own login logs.
Longer term, avoid exposing management interfaces for AI infrastructure directly to the internet. Put them behind a VPN or an allow-list, and keep a record of which internal tools are exposed. A recurring outside-in check can help you notice a service that is more reachable than you intended; Webforta Scan runs read-only scans of sites you verify, and the security headers checker can review the response headers of web interfaces you publish. Neither replaces patching GPUStack.
Timeline and sourcing
| Date | Event |
|---|---|
| Sep 30, 2026 | Advisory published in the gpustack/gpustack repository (per the GitHub advisory page) |
| Nov 4, 2026 | Published by the National Vulnerability Database |
| Nov 10, 2026 | Published to the GitHub Advisory Database, reviewed and last updated |
The advisory credits Austin Chu, Sohee Kim and Corban Villa (UC Berkeley security research project) for private reporting, and also an anonymous researcher who reported it independently and earlier. Only one independent publisher covers this story so far, and the technical details here come from the maintainers' advisory. You can also read the NVD entry for the record.
For comparable authentication flaws in other infrastructure tools, see our write-ups on the Kubero JWT auth bypass and the Kubeflow Pipelines upload authorization bypass.
Sources
This article is based on the public sources below. Details can change as vendors and researchers publish more; check the original advisories before acting on version numbers.
