Docker Model Runner argument injection enables container-to-host escapes
Original title: Hostile Inference: Exploiting Docker Model Runner for Container Escapes
ZDI says a default container can corrupt the Docker Desktop VM disk; non-default setups yield host code execution.
Trend Micro's Zero Day Initiative details ZDI-26-150 / CVE-2026-28400, an unauthenticated argument-injection flaw in Docker Model Runner, reported in November 2025. A container with default privileges can overwrite the Docker Desktop VM data disk, causing denial of service on macOS. Under non-default configurations, the same primitive becomes arbitrary code execution on the macOS or Windows host through Docker Extensions, plus further Seatbelt and WSL2 escapes.
Unconfirmed: Host code execution is described only under non-default configurations, and the post does not state a fix.
- The DMR API is unauthenticated and reachable at model-runner.docker.internal.
- _configure appends attacker JSON as llama-server command-line arguments.
- Inference runs on the host as a child of Docker Backend, not inside the VM.
- macOS uses a Seatbelt sandbox; Windows uses Job Objects.
- Docker Desktop for Windows containers and for Linux are out of scope.
Impact: VM disk corruption or host code execution
| Product | Affected |
|---|---|
| Docker Model Runner | — |
| Docker Desktop | macOS and Windows |
Vulnerabilities mentioned
All| CVE | Vulnerability | CVSS | EPSS | Signals | Published |
|---|---|---|---|---|---|
| CVE-2026-28400 | Docker Model Runner (DMR) is software used to manage, run, and deploy AI models using Docker | 7.5 | <1% |
From the original article
13 min read at thezdi.com
Local LLM inference is seeing adoption in default developer workflows, and the tooling that serves those models now ships inside products developers already trust. One of those is Docker Model Runner (DMR), the feature in Docker Desktop that pulls, runs, and serves models locally.
By Nitesh Surana · thezdi.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.