The AI Server That Believes Anyone Who Says 'I'm Home'

GPUStack CVE-2026-102895 hands attackers admin with one fake header; here's how to check and patch your clients' GPUStack servers.

Imagine a building whose front desk waves through anyone who says "I already work here." That is roughly how older GPUStack servers treated visitors. A request that claimed to come from the machine itself got full admin rights with no password, and attackers have already used that to run reverse shells on GPU hosts.

TL;DR

  • What happened: GPUStack before 2.2.0 gives unauthenticated admin to any request that looks like it came from 127.0.0.1. CISA added CVE-2026-102895 to KEV on 2026-11-09 after in-the-wild attacks.
  • Am I exposed? Any client GPUStack server below 2.2.0 that is reachable from the internet, or that sits behind a same-host reverse proxy, is exposed. GitHub's advisory lists both routes.
  • What to do: Upgrade to 2.2.2, which also fixes a SAML login bypass. Then hunt for rogue admin API keys and inference backends.
  • Impact: Attackers can mint admin keys, read cloud secrets and run code on every GPU worker the server controls.

What is GPUStack CVE-2026-102895?

CVE-2026-102895 is an authentication bypass in GPUStack, the open-source manager teams use to run AI models across a cluster of GPU machines. NVD rates it 9.8 (CVSS), the critical end of the scale, with no login and no user click needed.

The flaw sits in one function. According to the GitHub security advisory, get_current_user in gpustack/api/auth.py treats any loopback request as the built-in administrator whenever force_auth_localhost is false. False is the default.

That logic made sense for someone typing commands on the server itself. It fails in two ordinary situations:

  1. A reverse proxy on the same machine. If Nginx or Caddy forwards traffic to GPUStack from 127.0.0.1, every outside visitor looks local. No trick is needed: the attacker just visits.
  2. A forged header. An attacker sends X-GPUStack-Real-IP: 127.0.0.1 and GPUStack believes it.

Think of the header as a visitor badge the visitor prints at home. GPUStack read the badge and never checked who printed it.

Is CVE-2026-102895 being exploited?

Yes. The GitHub advisory says exploitation was observed in the wild against a GPUStack v2.1.1 deployment, where the attacker created custom inference backends that ran reverse shells. It credits Javier Sixto of ulex.fr with reporting the attacks. CISA's KEV entry gives federal agencies a due date of 2026-11-12, three days after listing, and marks ransomware use as unknown.

Nobody has published how many servers were hit, when attacks began or which route was used. SecurityOnline notes that install counts are private, and no indicators of compromise have been released.

Why does an AI server bug matter so much?

A GPUStack admin is not just a website admin; it controls a fleet of expensive machines. Per the advisory, an attacker with this access can change users and models, create admin API keys with scope ["*"] that work from any IP, and read worker credentials and cloud-provider secrets. The admin-only endpoint POST /v2/inference-backends/from-yaml then turns that into code running on worker hosts.

So one forged header can end with a stranger holding a client's cloud keys. That is the same pattern seen when a compromised AI gateway package leaked cloud credentials: AI tooling sits next to the most valuable secrets in the building.

Am I affected? How to check your clients' GPUStack servers

Your clients are affected if any GPUStack server runs a version below 2.2.0. For the related SAML flaw, CVE-2026-102894 (CVSS 9.1), anything from 0.7.1 up to 2.2.2 is affected. Work through these checks in order.

  1. Preserve evidence first. Before changing anything, copy the proxy logs, the GPUStack data volume and the container log, then hash them. Adjust the volume path if the client used a bind mount.
D=/root/ir-gpustack-$(date +%F); sudo mkdir -p "$D"
sudo cp -a /var/log/nginx /var/lib/docker/volumes/gpustack-data/_data "$D"/
for c in $(sudo docker ps -aq --filter ancestor=gpustack/gpustack); do sudo docker logs "$c" 2>&1 | sudo tee "$D/container-$c.log" >/dev/null; done
sudo find "$D" -type f -exec sha256sum {} + | sudo tee "$D.sha256"
  1. Find the version. This covers pip installs and every running GPUStack container on the host. Anything below 2.2.0 is vulnerable.
pip show gpustack 2>/dev/null | grep -i '^version'
for c in $(sudo docker ps -q --filter ancestor=gpustack/gpustack); do echo "$c"; sudo docker exec "$c" pip show gpustack | grep -i '^version'; done
  1. Test the bypass from outside, with permission. A 200 on either line means the server hands out admin access without a password. A 401 on both means this route is closed.
U=https://gpu.client.example
curl -s -o /dev/null -w 'plain: %{http_code}\n' "$U/v2/users"
curl -s -o /dev/null -w 'spoofed: %{http_code}\n' -H 'X-GPUStack-Real-IP: 127.0.0.1' "$U/v2/users"
  1. Hunt for rogue admin API keys. Using a known-good admin key, list every user's keys with full scope and look for names or creation dates nobody recognises. The user_id=* filter is what makes an admin see all users' keys, not just their own.
curl -s -H "Authorization: Bearer $ADMIN_KEY" "$U/v2/api-keys?user_id=*&perPage=100" | jq '.pagination, (.items[] | select(.scope|index("*")) | {name, user_name, created_at})'
  1. Hunt for shell-running inference backends. Any match is a backend that opens a connection out, which is how the observed attack worked.
curl -s -H "Authorization: Bearer $ADMIN_KEY" "$U/v2/inference-backends?perPage=100" | jq -c '.items[]' | grep -Ei 'dev/tcp|bash -i|nc -e|mkfifo|curl |wget '
  1. Check worker hosts for live shells. This shows only processes still running, so run it before rebooting anything.
ps -eo pid,ppid,user,comm,args --forest | grep -Ei -B2 -A3 'gpustack|vllm|/dev/tcp|bash -i'

How do I fix CVE-2026-102895?

Upgrade every GPUStack server to 2.2.2 or later, which closes both CVE-2026-102895 (fixed in 2.2.0) and the SAML bypass (fixed in 2.2.2). Then clean up.

  1. Preserve evidence using check 1 above, if you have not already.
  2. Upgrade with pip install -U 'gpustack>=2.2.2' or by pulling a 2.2.2+ image, then rerun check 2 and confirm the version reads 2.2.2 or higher.
  3. Rerun check 3 and confirm both lines return 401.
  4. Delete every unrecognised key and backend from checks 4 and 5, then rerun both and confirm only expected names remain.
  5. Rotate the cloud credentials the server stored (listed at GET /v2/cloud-credentials) in each cloud provider's console, then confirm a call with an old key is denied.

If a client cannot upgrade today, the advisory's stopgaps are: start GPUStack with --force-auth-localhost (or force_auth_localhost: true in the file passed to --config-file, per the CLI reference), and make the proxy strip incoming X-GPUStack-Real-IP headers. In Nginx, proxy_set_header X-GPUStack-Real-IP ""; drops it. The SAML advisory says no setting fixes that flaw, so block /auth/saml/callback at the proxy until 2.2.2 is in. Rerun check 3 afterwards; both lines must return 401.

Where KinoSec fits

The hard part for a provider is not patching GPUStack; it is knowing a client runs it at all. AI servers often get stood up by a data team on a spare box, put behind a quick reverse proxy and never added to the asset list.

That is where autonomous web app and API pentesting for your clients earns its keep. KinoSec maps each client's external surface, finds exposed management APIs like GPUStack's, and tests them the way an attacker would, with the client's permission, agreed scope and a report they can act on. It also shows how a single bypass chains into a bigger compromise across surfaces, which is exactly how a fake header became code on GPU workers here.

Frequently Asked Questions

Is the GPUStack vulnerability being exploited?

Yes. GitHub's advisory says exploitation was observed in the wild against a GPUStack v2.1.1 deployment, with custom inference backends running reverse shells, and credits Javier Sixto of ulex.fr with reporting it. CISA added CVE-2026-102895 to its Known Exploited Vulnerabilities catalog on 2026-11-09.

How do I check if my clients were hit by CVE-2026-102895?

First, preserve proxy logs and the GPUStack data volume. Then list every full-scope API key with GET /v2/api-keys?user_id=* and look for names you don't recognise. Finally, search inference backends for shell strings with jq -c '.items[]' | grep -Ei 'dev/tcp|bash -i'.

Which GPUStack version fixes the vulnerability?

GPUStack 2.2.0 fixes CVE-2026-102895, and 2.2.2 also fixes the SAML login bypass CVE-2026-102894. Upgrade to 2.2.2 or later. Until then, enable --force-auth-localhost, strip X-GPUStack-Real-IP at the proxy and block /auth/saml/callback.