Back to Blog

The patch for your AI server shipped in June; the advisory, in November

Close-up of a dual-fan graphics card resting on a desk

On 9 November CISA added CVE-2026-102895 to its catalogue of exploited vulnerabilities: an authentication bypass in GPUStack, the GPU cluster manager a lot of people have installed to serve models in house. A 9.8. No credentials, full administrator. That is not the uncomfortable part. The uncomfortable part is that the fix has been published since 18 June, in version 2.2.0, and that the notes for that release —thirteen thousand characters, thirty-six listed fixes— never once say "authentication", "security" or "localhost". The fix was there. The advisory took almost five months.

The rest of the calendar: the CVE record is published on 4 November; the project advisory and the entry into CISA's catalogue arrive on the 9th, with a remediation deadline for US federal agencies on the 12th, three days later. And in between, real exploitation against an installation running 2.1.1, released back in March.

The standard nginx recipe is already enough

The bug fits in three lines of Python. In 2.1.1, get_current_user ends like this: if nobody has identified themselves and the source address looks like 127.0.0.1, you are handed the first administrator user in the table. Not a local user with trimmed permissions: the administrator. The condition that prevents it does exist —force_auth_localhost— and in config.py it is declared as bool = False. Off by default. The NIST record says it plainly: "treated a loopback request source as the built-in administrator when force_auth_localhost was false, which was the default".

That leaves two ways to look local, and the first is what turns this into a real problem. Nothing needs forging: the project advisory describes "the canonical self-hosted shape", GPUStack behind a same-host reverse proxy with proxy_pass http://127.0.0.1:<port>, and in that shape "every remote request arrives at GPUStack with TCP peer 127.0.0.1, so the loopback auto-admin branch fires. No header or credential is required." In other words: the recipe everybody copies and pastes to publish an internal service is exactly the one that opens it. The second route is the X-GPUStack-Real-IP header: in 2.1.1, when gateway_mode is set to embedded, the function that works out the source returns that header's value exactly as it arrived. The client writes it. The server believes it.

The advisory writes the root cause itself, in a sentence that applies far beyond this product: "Trusting the network source of a request (socket peer, or a client-controllable header) as an authentication signal." Treating where a request comes from as if it were a credential. A few days ago we wrote about another authentication bypass, the OpenSSH one on musl, where the hole was also not where everyone was looking. Here the cause is older still: a development convenience nobody revisited when the service went into production.

What whoever gets in walks away with

CISA gives the list in its entry, written in the conditional —"could allow"— as these things are: user and model changes, creation of persistent administrator API keys, disclosure of worker credentials and cloud-provider secrets, and code execution through POST /v2/inference-backends/from-yaml. The project advisory adds the detail that matters about those keys: scope ["*"] and "valid from any source IP thereafter". And it confirms this is not theory: "This vulnerability has been observed exploited in the wild against a v2.1.1 deployment (attacker-created custom backends running reverse shells)." Somebody got in and set up their own inference backend that opened connections outbound.

What upgrading does not fix

This is the part that gets skipped most often. Put the two previous sentences together: an administrator key created by the intruder is a row in your database, with full scope and valid from any IP thereafter. Upgrading to 2.3.1 closes the door. It does not delete the copy of the key that was cut while it stood open. If that machine has been reachable from outside, upgrading is the first of three steps: upgrade, revoke one by one every API key you do not recognise, and rotate the provider secrets it was holding, because they are not encrypted against you: they are there for the service to use.

Three signals, and none of them says the same thing

While writing this we looked up the EPSS score for this CVE, the number many tools use to sort the patching queue. Dated 14 November —five days after it entered the exploited catalogue— it returns 0.00517, percentile 0.4456. Translated: half a per cent probability, and below the median of all scored CVEs. The same CVE CISA says is being exploited.

Signal What it says about this CVE Which question it answers
CVSS 3.1 9.8 — critical If it happens, how bad is it?
EPSS (14-nov) 0.5 % — 44.6th percentile Will a lot of exploitation be seen across the internet?
CISA KEV (9 Nov) Exploited; federal deadline, 12 Nov Has it already happened, at least once?

EPSS is not broken: it answers a different question, and FIRST writes it into its own FAQ —"EPSS scores reflect exploitation probability across the global observed population, not your specific environment"—. It is the probability that, within the next thirty days, one of their data partners' sensors will see exploitation activity: a probability computed over the global population those sensors observe, not over your building. GPUStack is popular among people building AI in house —a little over six thousand stars on GitHub— but it is not Apache: there are not hundreds of thousands of instances making noise out there. A niche product can be actively exploited against your fifty installs and still score a mediocre percentile, because fifty are not visible from that altitude.

From which comes a practical rule that costs nothing: the percentile is for sorting what you have in common with the world —the browser, the web server, the hypervisor— and it is no good for deciding about the odd thing you built yourself. For the odd thing, something else decides: whether it is exposed, what credentials it holds and who watches it. If your patching criterion is "above such-and-such percentile", this 9.8 stays in the queue next to the building's air conditioning.

The question nobody asks: whose machine is that?

Almost every AI box we have seen in small and mid-sized companies has the same biography. Someone in development asked for a GPU for a trial. It was bought quickly, because it was for research and nobody classed it as infrastructure. It was published behind the proxy that already existed so the team could reach it from home. The model providers' keys were loaded into it to compare costs. And there it still sits, a year later, serving things that are no longer a trial. It is not in the inventory, it has no update window and, above all, it has no owner: when you ask whose it is, the answer is a person's name, not a team's, and that person stopped looking at it long ago.

We have one of those ourselves and we are not going to lecture from the outside: we use self-hosted n8n for internal automation. What we did with it has no technical merit whatsoever, and it is exactly what is missing on the machines this post is about: it is in the inventory along with the rest of the infrastructure, it has an owner, and it has an update window, like any other production service. It is boring. That is why it works: the day its own CVE lands, nobody will have to find out who installed it.

What to check today, in this order

  1. Whether it exists. Do not look at the inventory: if it were in the inventory you would not be making this list. Look at the hardware purchases of the last eighteen months and ask who requested a GPU.
  2. Which version it runs. Fixed in 2.2.0 (18 June); the current branch is 2.3.1, from 11 November. Anything below 2.2.0 is vulnerable.
  3. Whether a reverse proxy sits in front on the same host. That is the condition that turns "it only listens locally" into "anyone is an administrator", with no header and no credential.
  4. The advisory's own two mitigations, if you cannot upgrade today: force_auth_localhost=true, and have the edge strip any client-supplied X-GPUStack-Real-IP header.
  5. If it was exposed: revoke API keys and rotate provider secrets before calling the incident closed. Upgrading does not delete them.

If you do not run GPUStack, this is not about GPUStack

The pattern —"if it comes from local, you are the administrator", plus a proxy in front— is not exclusive to this product. It lives in admin panels, in monitoring interfaces, in internal dashboards and in more than one tool you set up in twenty minutes because it was just for testing. The question that works for all of them is the same: what decides here that a request is local, a fact of the connection or a claim made by the client? The fixed version of GPUStack writes it into a code comment, and it is the best summary of the criterion we have read in months: what you accept has to be "a fact about the connection rather than a client-supplied claim".

And the honest caveat, because not everything deserves the same ceremony: the Raspberry Pi reading a thermometer in the warehouse needs neither a formal inventory entry nor an update window, and convening a committee for it is the fastest way to make sure nobody takes the process seriously. The machines that do need it are the ones that hold credentials to other things. An AI box almost always does: that is why it was built, to talk to model providers and to the company's data. It is the same snag we wrote about regarding pinning the version of an MCP server: whatever you decide to treat as a "testing tool" ends up with the permissions of a production service.

At everyWAN we do not sell GPUStack or any product from this ecosystem, so the only thing we take from here is the criterion. If you want someone to look at what you have powered on, on which version and exposed to what, that is cybersecurity work on the boring side: inventory, exposure and who reads the logs. And if what is on the table is building AI in house without it ending up as a box with no owner, that is exactly the automation and AI conversation, with the maintenance decided on day one rather than on advisory day.

Sources (verified): description of the flaw, force_auth_localhost defaulting to false and the record's publication date (4 Nov 2026) — NVD, CVE-2026-102895; the two exploitation routes, the sentence about the canonical shape with proxy_pass, the scope of the API keys, the root cause, the mitigations and the exploitation observed against a 2.1.1 deployment — advisory GHSA-gj9m-f8ph-pmhr (published 9 Nov 2026, CVSS 3.1 9.8); catalogue entry date (9 Nov), federal deadline (12 Nov) and the list of what the flaw allows — CISA KEV catalogue, version 2026.11.13; EPSS 0.00517 and percentile 0.4456 dated 14 Nov 2026 — EPSS API and FAQ (FIRST); release dates for 2.1.1 (26 Mar 2026), 2.2.0 (18 Jun 2026) and 2.3.1 (11 Nov 2026), the contents of the 2.2.0 notes and the code of auth.py and config.py at each tag — GPUStack repository. That the 2.2.0 notes mention neither authentication, security nor localhost is our own check against that release's published text. The "AI box" biographies are patterns we see in the field, not one specific customer.

Who owns your company's AI box?

If the answer is a person's name rather than a team's, or it takes more than ten seconds to arrive, that machine has no update window. We will look at it with you: what is powered on, on which version and exposed to what.

Talk to everyWAN

Tags:

Share:

Subscribe to our newsletter

To receive IT stories, everyWAN news and exclusive subscriber offers, sign up to our mailing list

Minorisa de Sistemas Informaticos y Gestión S.L. © 2026
everyWAN
everyWAN