thezdi.comNitesh Surana
Researchers Show Docker Model Runner Can Escape Containers
Part of the CVE-2026-28400 story · since 18 Nov
- Wed 18 Nov, 20:30 UTC. thezdi.com, first report. Researchers Show Docker Model Runner Can Escape Containers. This article, 1st of 1.
Trend Micro’s Zero Day Initiative researchers publish a detailed chain showing how Docker Desktop’s Local LLM inference feature, Docker Model Runner (DMR), can be abused to escape from a stock container to the host.
The investigation revisits a previously disclosed denial-of-service bug (CVE-2026-28400, ZDI-26-150) that allowed a container with default privileges to affect the Docker Desktop VM’s data disk on macOS, and extends it to demonstrate arbitrary code execution on both macOS and Windows under non-default configurations.
The core finding is that DMR runs the host-side llama-server process with attacker-controlled arguments, and that, via the unauthenticated DMR REST API exposed at model-runner.docker[.]internal, a container can influence the host-bound invocation of the inference server.
The researchers map a full attack chain: beginning with an argument-injection primitive in the _configure endpoint, then gaining host file-read/write footholds through three llama-server flags that touch the host filesystem (--grammar-file, --chat-template-file, --log-file). On macOS, a permissive Seatbelt sandbox can leak and misuse host paths (including the Docker data disk at vms/0/data/Docker[.]raw) to overwrite logs, while on Windows a Job Object sandbox permits host access more broadly.
By combining this with payloads in GGUF models and leveraging Docker Extensions (via ddClient.docker.cli[.]exec and ddClient.extension.host.cli[.]exec), the chain pivots from partial host-write control to arbitrary code execution on the host. The PoC and proof points cover how the extension canvas can be leveraged to drop and execute a host binary, with notable username leakage steps in Windows via UNC paths or error logs.
ZDI notes subsequent vendor responses and hardening steps, including fixes that constrained flags, implemented path-safety checks, and restricted back-end allowlists, culminating in patched releases in 2026.
The piece ties these technical steps to broader disclosure timelines and vendor patching, emphasising that local inference surfaces attack paths not automatically contained by container isolation or extension sandboxing. The report attributes specific exploit steps and mitigations to ZDI researchers and Docker's development team, with cross-platform implications highlighted for macOS Seatbelt and Windows Job Objects plus Docker Extensions.
It also points to related analyses from prior advisories and companion posts on VM-to-host escapes via Docker Extensions and WSL2, reinforcing the need for careful review of local-AI supply chains and extension trust models. The article preserves the CVEs and version numbers (e.g., CVE-2026-28400; DMR v1.0.16; Docker Desktop 4.61.0+), and records the vendor’s patch cadence and the coordinated release timeline.
How the story developed
1 article from 1 outlet since 18 Nov. Still developing; the latest report was 6 hours ago. This was the first report.
- Wed 18 Nov, 20:30 UTC. thezdi.com, first report. Researchers Show Docker Model Runner Can Escape Containers. This article, 1st of 1.
- CVE not in KEV
- No CVE
- Larger: an outlet’s first report
- Coverage over time
Show all 1 articles as a listHide the list
- Wed 18 NovDay 01
- Researchers Show Docker Model Runner Can Escape Containers20:30 · thezdi.com · First report · This article, 1st of 1