CVE-2026-33990
Docker Model Runner token exchange can reach and leak internal services
Docker · Docker Model Runner OCI registry client
First documented disclosure · Review checkpoint
Summary
An attacker-controlled model registry can redirect token requests to internal services and return their data.
Severity and exploitation evidence
Medium · 6.8/10 · CVSS 4.0. Severity source.
Exploitation not established: The reviewed advisory and public fix establish an SSRF weakness; they do not establish malicious exploitation or confirmed victims.
Technical severity is not a measure of observed attacks or of risk to your deployment. Attempts, demonstrations, and successful compromise retain the qualifications in the source.
How the weakness works
The registry WWW-Authenticate realm URL is followed without scheme, host or IP validation. Internal response bodies can be reflected to the caller or relayed to the registry in Authorization: Bearer.
Conditions required
A malicious OCI registry must be used for a model pull. An attacker can trigger that through accessible Model Runner API access, or a user can pull the attacker-controlled model.
Potential impact
Server-side GET requests from the host to internal services and disclosure of their response data. Arbitrary host code execution is not established by this advisory.
Affected versions and configurations
Docker Model Runner before 1.1.25.
Fix and mitigation status
Model Runner 1.1.25; Docker Desktop 4.67.0 includes the fix.
Practical next steps
Upgrade to a supported release containing the fix. Restrict model registries and Model Runner API exposure. Enhanced Container Isolation blocks the container access path, but the advisory retains a caveat for localhost TCP exposure.
Operating lesson
Treat registry authentication challenges as untrusted redirects, and validate their destinations before making host-network requests.
Evidence and limits
The public security-specific PR opened and merged March 24, before the March 30 GHSA and April 1 CNA record; March 24 is therefore used for disclosure. Model Runner 1.1.25 shipped March 24 and Desktop 4.67.0 on March 30. The CNA scores 6.8 Medium under CVSS 4.0 (local, low privileges, no interaction); the current GHSA scores 7.1 High under CVSS 4.0 (network, no privileges, passive interaction). These are source-specific assumptions, not interchangeable scores.
Sources
- GitHub CNA record
Primary source ·
- Docker maintainer advisory
Primary source ·
- Original public SSRF fix PR
Primary source ·
- Model Runner 1.1.25 release
Primary source ·
- Desktop 4.67.0 release notes
Primary source ·