Skip to main content
Menu
CVEs

CVE

LLMKube arbitrary host file read via Model hostPath injection

Affects LLMKube before 0.9.0 when model caching is enabled, which is the default. Users with create or update rights on Model and InferenceService resources can read files outside their namespace. Fixed in 0.9.0.

CVE-2026-68903Published Updated Summary written by a model

7.7HighCNA
Exploited
No known exploitation
Fix
Fixed in 0.9.0
Who is exposed
Specific setups only

Do first: Check whether your setup matches what the bug needs. Model caching enabled (modelCache.enabled), which is the default Controller accepts arbitrary absolute paths and file:// URIs in Model.spec.source Attacker has create or update rights on Model and InferenceService resources in a namespace

Summary

LLMKube, a Kubernetes operator for self-hosted large language model inference, can be reached over the network by a user with low-privilege access to the cluster and no user interaction. That user can create or update Model and InferenceService resources to make the operator copy arbitrary files from the controller pod or a cluster node into model storage, exposing service-account credentials and other secrets beyond their namespace.

Original description

LLMKube is a Kubernetes operator for self-hosted large language model inference across heterogeneous hardware. Prior to 0.9.0, the ModelReconciler accepts arbitrary absolute paths and file:// URIs in Model.spec.source without restricting them to approved roots. A namespace user who can create or update Model and InferenceService resources can cause copyLocalModel to open an attacker-selected file in the controller pod or cause a HostPathVolumeSource to mount an attacker-selected node file when modelCache.enabled is enabled, which is the default. The selected file is copied into model storage and can expose controller service-account credentials or other node secrets outside the attacker's namespace. This issue is fixed in version 0.9.0.

Who should check

Teams running LLMKube 0.8.28 or earlier on Kubernetes with default Helm values, and anyone letting namespace users write Model or InferenceService resources

“LLMKube 0.8.27 or 0.8.28 installed with default Helm values”

github.com

cve.now Index

16Pcve.now Index 16 of 100 · methodology v3 · provisional
  • Exploitation?/40Unknown. No CISA SSVC exploitation rating yet
  • Exposure8/25Specific setups only, reachable over the network
  • Reach?/20Unknown. No tracked technology or package matched
  • Severity7.7/10CVSS 7.7
  • No patch0/5No signal

What would change the score

  • Would score about 66 if added to CISA KEV
  • Would score about 61 if CISA's SSVC rates exploitation active
  • Would score about 23 if CISA's SSVC rates it automatable

What to do

  • Check whether your setup matches what the bug needs.

    Model caching enabled (modelCache.enabled), which is the default Controller accepts arbitrary absolute paths and file:// URIs in Model.spec.source Attacker has create or update rights on Model and InferenceService resources in a namespace
  • Upgrade to a fixed version: 0.9.0.

Check your systems with an agent

Paste this into a coding or ops agent. It checks your systems read-only and reports a table per asset.

Check whether my systems are affected by CVE-2026-68903 (LLMKube: Arbitrary Host File Read via Model.spec.source hostPath Injection), using the cve.now data below.

Rules: work read-only. Do not run exploit code, proof-of-concept scripts or scanners that send attack payloads, and do not change any configuration, package, image or deployment. Report findings only.

What cve.now has for CVE-2026-68903 (entry last updated 2026-11-18):
Affected:
- defilantech LLMKube: fixed in 0.9.0
Fixed versions: 0.9.0.
Conditions (all must be true for the bug to be hit):
- Model caching enabled (modelCache.enabled), which is the default (source: https://github.com/defilantech/LLMKube/security/advisories/GHSA-jw3m-8q7m-f35r)
- Controller accepts arbitrary absolute paths and file:// URIs in Model.spec.source (source: https://github.com/defilantech/LLMKube/security/advisories/GHSA-jw3m-8q7m-f35r)
- Attacker has create or update rights on Model and InferenceService resources in a namespace (source: https://github.com/defilantech/LLMKube/security/advisories/GHSA-jw3m-8q7m-f35r)
Setup: the bug is reachable in a default setup; no special feature or setting is needed.
Attacker needs: network access, low-privilege login, no user action.
CISA KEV: CVE-2026-68903 is not listed as of cve.now's last read.
This KEV status is for CVE-2026-68903 only. If you also check another CVE, look up its own KEV status; never assume it from this one.

How to check:
- Installed software: check what is actually running, not what a repo declares. Read the version from each host, container image and Kubernetes node (the binary's version output, the OS package database, image layers), and check whether the conditions above hold there. Distribution packages may carry a backported fix under an older version number; confirm against the distribution's own advisory.

Verify against the linked advisories before concluding, and cite the date you read each one:
- https://github.com/defilantech/LLMKube/security/advisories/GHSA-jw3m-8q7m-f35r
- https://github.com/defilantech/LLMKube/pull/981
- https://github.com/defilantech/LLMKube/commit/ec2647fd4557962a971a1bc9435790b8e77ce961
- https://github.com/defilantech/LLMKube/releases/tag/v0.9.0
Do not conclude an asset is unaffected unless you checked it and the evidence covers both its version and the conditions above. If you could not check something, mark it unknown and say what is missing.

Report a table with one row per asset and these columns: Asset | Version or build | Status (affected, fixed, not applicable, unknown) | Compensating controls | Exposure (internet-facing, internal, none) | Evidence | Next action.

Entry: https://cve.now/CVE-2026-68903
For follow-up questions, use the cve.now MCP server at https://mcp.cve.now/mcp (no authentication), for example execute with name cve.get and arguments {"id": "CVE-2026-68903"}.

Timeline

  1. CVE published
  2. First seen by cve.now

Affected versions

  • defilantech LLMKube

    fixed in 0.9.0

Evidence 6

KindSource
Vendor advisory 1Establishes: Affected and fixed versions.
AdvisoryGitHub · defilantech/LLMKube · Advisory GHSA-jw3m-8q7m-f35r
CVE record 2Establishes: The identifier and description.
RecordCVE List · CVE-2026-68903
RecordNVD · CVE-2026-68903
Other links 3
RepositoryGitHub · defilantech/LLMKube · Pull request #981
RepositoryGitHub · defilantech/LLMKube · Commit ec2647fd
RepositoryGitHub · defilantech/LLMKube · Release v0.9.0

Corrections

None so far.

Something wrong on this page? Report an error. Every fix is listed here.

Questions about CVE-2026-68903

Is CVE-2026-68903 being exploited?

It is not in CISA's Known Exploited Vulnerabilities catalog, and CISA's SSVC assessment does not mark exploitation as active. That does not mean it has not been attacked: vendors and researchers can report attacks before CISA lists them. No public exploit code is tracked.Source: cve.now feed data

What does this mean?

A namespace user who can create or update Model and InferenceService resources can set Model.spec.source to an arbitrary absolute path or file:// URI. The controller then reads files it can access, including its service-account token and node secrets, from outside the attacker's namespace.Source: github.com

Which versions are affected by CVE-2026-68903?

LLMKube: before 0.9.0.Source: cve.now feed data

Am I affected, and how do I check?

Versions before 0.9.0 are affected, and the exposure requires model caching (modelCache.enabled), which is enabled by default in the CLI and the official Helm chart. To check, look at the installed LLMKube version and whether modelCache.enabled is set in the Helm values.Source: github.com

How do I fix CVE-2026-68903?

Upgrade LLMKube to 0.9.0 or later, which fixes this vulnerability.Source: github.com

More CVEs