CVE-2023-31036 is a high severity (CVSS 7.5) relative path traversal in NVIDIA Triton Inference Server for Linux and Windows. It affects only deployments launched with explicit model control, and NVIDIA's fix is Triton 2.40 or later, in the release branch since 4 December 2023.
What's affected
NVIDIA's bulletin lists all versions of Triton Inference Server on Linux and Windows prior to 2.40 as affected, and 2.40 as the updated version. The issue affects only nondefault deployments that enable dynamic model loading through the model control APIs with the explicit model control option. Deployments with default settings are not affected.
NVIDIA says the fixed code was in the development branch on 10 November 2023 and in the release branch on 4 December 2023. The update adds the ability to restrict the HTTP endpoint of the model load API and prevents the model load API from accessing directories outside the model directory.
How to check if you're exposed
Check how your server is launched. The option in question is explicit model control, set on the command line. On a host, inspect running processes:
ps aux | grep tritonserver
For a container, view its command and the image tag:
docker inspect --format '{{.Config.Cmd}} {{.Config.Image}}' CONTAINER
Map the image tag to the Triton server version using NVIDIA's release notes, since container tags are not the same as server versions. If the launch options include explicit model control and the version is below 2.40, you are exposed. Safeguard's free in-browser scan at https://gold.safeguard.sh/scan can also flag known issues in container dependency lists.
What to do now
- Find every Triton Inference Server deployment and whether it runs with explicit model control.
- Upgrade to 2.40 or later.
- If you need explicit model control, restrict the HTTP endpoint of the model load API to trusted administrators, as the updated software allows.
- Read NVIDIA's Secure Deployment Considerations Guide for Triton, which the bulletin links.
Why it matters
The record lists CWE-23, relative path traversal, grouped under CWE-22. A name that should point inside a model folder includes parts that move outside it, and the server follows. NVIDIA says the impact may include code execution, denial of service, privilege escalation, information disclosure and data tampering.
The flaw matters because inference servers load model files, and a model is a kind of code. Anyone able to call the load API with a path outside the intended folder gets the server to open files of their choosing. The attack complexity is high and the vector needs low privileges, and the narrow set of deployments with explicit control narrows exposure. NVIDIA's reporter credit goes to l1k3beef at tencent-zhuquelab. The CVE is not on the CISA KEV list.
Try it yourself
Click through it at your own pace, or open it full screen.
Quick check
Am I affected if I use default settings?
No. NVIDIA says deployments using default settings are not affected.
What does the fix change?
It lets you restrict the model load API endpoint and stops it from reaching directories outside the model directory.
Sources
Last reviewed: 19 October 2026.