Opens in a new tab
News

GitLab’s critical patch: check the AI Gateway’s own version

In brief

A successful GitLab application upgrade can leave a separate AI Gateway unchanged. Follow the critical advisory to the component and operator responsible for its patch.

3 min read

Is this the right starting point?
Before you read
Basic Docker or Helm deployment and runtime-version inspection skills.
What you will take away
Identify whether a self-hosted GitLab AI Gateway needs the patch and verify the actual running gateway release.
Sources
A coral architectural partition has one silver panel, with pale ribbons ending at the wall.
On this page3 sections

The GitLab application shows the version the administrator expected. The upgrade worked. But in this hypothetical installation, a separate AI Gateway container is still running 19.4.0. The version on the application screen cannot tell you whether that other component has been patched.

GitLab’s critical AI Gateway advisory is about the gateway. Under the described conditions, an authenticated Duo Agent Platform user could submit a crafted custom-flow configuration that escapes prompt-template restrictions and executes commands there. The listed fixed releases are 19.2.4, 19.3.2 and 19.4.1.

Find the operator responsible for this gateway

First establish who runs this gateway. GitLab says it has fixed the gateways it hosts, including those serving GitLab.com, Dedicated and Self-Managed customers. Running the main application yourself does not necessarily mean you host the gateway. That distinction tells you which operator needs to act.

For a customer-hosted AI Gateway, compare its release with these affected ranges:

Affected AI Gateway release rangeListed fixed release
18.1.6 up to, but excluding, 19.2.419.2.4
19.3 up to, but excluding, 19.3.219.3.2
19.4 up to, but excluding, 19.4.119.4.1

The CVE-2026-90970 record was published October 2. GitLab rates the issue critical, with a CVSS 3.1 score of 9.9. The flaw concerns how the server handles a crafted flow configuration; it is not simply a language model agreeing to ignore an instruction.

Patch scope starts with who hosts the AI Gateway. GitLab-hosted gateways are already fixed; customer-hosted gateways need their actual running release checked against the affected ranges.
GitLab’s advisory makes gateway hosting the first branch in the patch decision. This diagram does not reproduce an exploit or describe a customer deployment.

Read the image that is actually serving requests

The application and gateway have their own running images. Inspect the gateway’s image rather than inferring it from the application. GitLab’s installation guidance covers Docker and Helm and calls for explicit stable tags aligned with the GitLab minor release.

In our hypothetical 19.4.0 case, the listed fix is 19.4.1. An older installation needs more care: this table lists no patch below the 19.2 line. Resolve the supported upgrade path with GitLab instead of assuming a newer gateway minor release can be dropped beside an older application.

The bounded-action approach keeps the change attached to that running component: identify it, choose a compatible corrected image and inspect it afterward. A revised tag in a deployment file expresses what should run. An old replica still accepting requests tells you what some users actually reach.

Finish the change at the running component

For Docker, inspect the gateway container and its image. For Helm, compare the release settings with the images in the running gateway pods and their rollout state. Follow the procedure for that deployment method, then check access to the AI Gateway and Agent Platform.

That establishes the running version and basic continuity, not the absence of an earlier compromise. It is not an exploit reproduction either. The CVE record’s October 2 CISA assessment listed exploitation as “none”; that dated observation is not a permanent guarantee. The sources reviewed here do not establish active exploitation.

The completed patch record should name the gateway’s operator, the image that served requests and the supported fixed release now running. Back at the administrator’s screen, the successful application upgrade was real. Checking the separate gateway is what finishes this particular repair.

Sources & context

Sources linked in this article. Read alongside the author’s analysis; a citation does not independently verify a publisher’s claims.

Report an error or outdated detail

Keep reading

Where to go next