Some software decides who you are by where your request appears to come from. GPUStack, an open-source cluster manager that teams use to run AI models across pools of GPUs, did that by default before version 2.2.0. A request that seemed to come from the machine itself was treated as the built-in administrator, with no password or API key. Attackers have now abused that behavior against a real deployment.

On Nov. 9, the Cybersecurity and Infrastructure Security Agency added CVE-2026-102895 to its Known Exploited Vulnerabilities (KEV) catalog. It listed the bug as a GPUStack "Improper Authentication Vulnerability" and set the remediation deadline for Nov. 12, three days after the listing. The CVSS 3.1 score is 9.8 out of 10, rated critical, because the attack works over the network, is easy to pull off, and needs no account and no help from a user.

Patching takes little effort, because the advisory and fix have been public since July. The harder part is working out what an attacker may already have done with admin access.

How "Localhost Means Admin" Went Wrong

The National Vulnerability Database record places the flaw in a function called get_current_user in GPUStack's gpustack/api/auth.py. In versions before 2.2.0, when a setting named force_auth_localhost was false, which was the default, the function treated any request from a loopback address (127.0.0.1, the machine talking to itself) as coming from the built-in admin.

Trusting localhost is a common shortcut. It assumes that anyone who can send traffic from inside the box already controls it. That assumption only holds if the server knows the real origin of a request. According to the GitHub security advisory, GPUStack's client-IP logic could be misled in two ways:

  • A reverse proxy on the same host. Many deployments put nginx, Caddy or a similar proxy in front of the application on the same machine. The proxy forwards every outside request, so to GPUStack each connection comes from 127.0.0.1. With localhost trusted, anyone on the internet who reached the proxy was effectively the admin.
  • A forged header. The NVD record says the code that works out the client's IP address would accept a loopback value placed in the X-GPUStack-Real-IP header by the client itself. A remote caller could simply claim to be local.

Either way, the result is the same: full administrative control with no authentication. The vulnerability is tagged CWE-287 (improper authentication), CWE-290 (authentication bypass by spoofing) and CWE-346 (origin validation error), which together describe the bug fairly precisely.

What an Attacker Gets

The NVD description lists a wide range of impact. An unauthenticated attacker could:

  • change users and models on the cluster;
  • create a persistent admin API key, which keeps working after the original hole is closed;
  • read worker credentials and cloud-provider secrets stored in GPUStack;
  • run code on worker hosts by registering an inference backend through the POST /v2/inference-backends/from-yaml endpoint.

The last item matters most. An "inference backend" is the runtime GPUStack starts to serve a model. If an attacker can define a custom backend, they can decide what runs on the GPU machines. The GitHub advisory says that is what happened: it reports exploitation in the wild in which attackers created custom backends that ran reverse shells, giving them interactive remote access to the hosts.

NVD adds one more detail: "Exploitation was observed in the wild against a 2.1.1 deployment." The sources do not name the victim, the attackers or how widespread the activity is. CISA lists ransomware use as "Unknown."

A Months-Old Fix and a Three-Day Clock

The timeline helps explain the short deadline. The GitHub advisory, GHSA-gj9m-f8ph-pmhr, was published on July 31, 2026. It rated the issue critical, listed every pip release of gpustack below 2.2.0 as affected, and named 2.2.0 as the patched version. NVD published its record on Nov. 4 and lists it as analyzed. CISA's KEV entry followed on Nov. 9.

By the time the bug reached the KEV catalog, administrators had had more than three months since the advisory to upgrade, which may help explain the short window; the sources do not say why CISA chose it. The catalog entry cites Binding Operational Directive 26-04 and marks the vulnerability for forensic triage ("forensicTriage: Yes"), which points to checking for compromise as well as installing the patch.

According to the advisory, the fix removes the automatic admin path for loopback requests and tightens how GPUStack handles client IP addresses and forwarding headers.

Two Other Bugs in the Same Batch

CISA added several other entries on Nov. 9 with the same Nov. 12 deadline, including these two:

  • CVE-2026-116779, SignalWire FreeSWITCH: a code-injection flaw in the open-source telephony platform. The KEV entry says it can expose configuration and credentials, give an attacker control over calls, and allow command execution.
  • CVE-2026-123218, IBM Aspera Faspex: an incorrect-authorization flaw in IBM's file-exchange product that could let an authenticated user escalate privileges.

A few days earlier, on Nov. 6, the catalog added SonicWall CVE-2026-19868, which ZeroDayFeed has covered separately and whose deadline is Nov. 20. If your organization tracks KEV deadlines, note that the Nov. 9 additions all fall due on Nov. 12.

What It Means for Defenders

Who is affected: anyone running GPUStack below 2.2.0, especially with the management API reachable from outside the host. That includes deployments behind a reverse proxy on the same machine, which sounds safer but was the clearest path to the bypass. Note that the observed exploitation hit version 2.1.1, so being only a version behind is not a reason to relax.

Patch: upgrade to GPUStack 2.2.0 or later. If you cannot upgrade immediately, the advisory lists workarounds: set force_auth_localhost to true, make sure your reverse proxy or gateway cannot cause external traffic to look like loopback, strip any client-supplied X-GPUStack-Real-IP header at the edge, and keep the API off untrusted networks. Upgrading is the remedy both the advisory and NVD point to.

Assume compromise if you were exposed. Admin access could have been used to set up lasting access, so upgrading does not undo what an attacker may already have done. Based on the advisory and NVD descriptions, these steps are ZeroDayFeed's recommendation (the advisory itself prescribes the upgrade and workarounds above):

  • Rotate admin API keys. A key created during an intrusion keeps working after the upgrade. Review every API key, revoke any you can't account for, and reissue the rest.
  • Rotate worker and cloud-provider secrets. NVD lists disclosure of these secrets as a direct impact. If your GPUStack instance had credentials for a cloud account, treat them as exposed and check that account for activity you don't recognize, such as new instances, new access keys or unusual spending on GPU capacity.
  • Review custom inference backends. The reported attacks ran their payloads through custom backends. Compare every backend definition against what your team actually deployed, and investigate any worker host that ran one you can't explain.

Check users and models as well. NVD notes that attackers could modify users and models. An admin account you don't recognize, or a model you didn't deploy, is a warning sign.

Look at the network path. If a reverse proxy sits in front of GPUStack, check your proxy and application logs for admin actions you can't tie to a staff member. Look in particular for requests that carried a client-supplied X-GPUStack-Real-IP header. More broadly, a GPU cluster manager rarely needs to be reachable from the public internet. Limiting access to a VPN or management network reduces the impact of the next authentication bug.

For federal contractors and CMMC shops: AI infrastructure set up quickly by a data-science team often never makes it onto the asset inventory. If anyone in your organization runs GPUStack, confirm it is on the list, scanned and patched before the Nov. 12 deadline.

Sources