Docker Model Runner Flaw Lets Containers Break Out Onto Developer Machines
A vulnerability in Docker Model Runner (DMR), the local AI inference feature built into Docker Desktop, allows an ordinary Linux container to reach across the container boundary and interfere with the host machine. The issue is tracked as CVE-2026-28400 and as advisory ZDI-26-150, and was reported to the vendor in November 2025. In its baseline form it is an unauthenticated argument-injection flaw that a container created with no extra capabilities, flags, mounts or privileges can trigger, producing a denial of service on macOS by overwriting and corrupting the data disk of the Docker Desktop LinuxKit VM. Follow-up analysis showed that under non-default configurations the same primitive can be escalated to arbitrary code execution on the host, on both macOS and Windows. Docker Desktop for Linux, and Docker Desktop for Windows running Windows containers, are outside the scope of this research.
The weakness sits in the DMR REST API, which is unauthenticated and reachable from any container on the internal DNS name model-runner.docker.internal over plain HTTP. Its /engines/_configure endpoint, backed by the ConfigureRequest structure, accepts two JSON properties, runtime-flags and raw-runtime-flags, whose values are appended directly to the command line of the llama.cpp llama-server process used for inference. Values supplied in runtime-flags are taken as-is, while raw-runtime-flags is tokenised through shellwords.Parse, and the resulting configuration is persisted per model and per mode so that it is applied the next time the runner starts. The crucial design detail is that on macOS and Windows the inference process does not run inside the Linux VM at all; it runs on the host as a child of the Docker Backend daemon (com.docker.backend or com.docker.backend.exe) with the privileges of the user running Docker Desktop. Injected flags therefore translate into file read and write access on the host filesystem, and with partial control over what gets written that constrained write was turned into host code execution by way of Docker Extensions. Two additional escapes were identified alongside the main chain: one arising from an overly permissive Seatbelt sandbox configuration on macOS, and one that pivots the WSL2 VM-to-host boundary, again through Docker Extensions.
The significance here is the trust placed in an internal, unauthenticated service. Anything running in a container on a developer laptop, including a dependency pulled from a public registry, a CI job, or a model treated as untrusted input, can speak to model-runner.docker.internal without credentials and without needing a privileged container. Developer workstations are high-value targets because they typically hold source code, signing material, cloud tokens and SSH keys, so a container-to-host escape on these machines has a far larger blast radius than the single application being tested. The sandboxes intended to contain the inference process, Seatbelt on macOS and Job Objects on Windows, are not equivalent boundaries, and that asymmetry carries much of the risk on the Windows side. On current exploitation status, FIRST EPSS places CVE-2026-28400 at a 0.3% probability of exploitation in the next 30 days, and there is no indication in the available evidence of exploitation in the wild; the flaw was disclosed through coordinated disclosure rather than discovered in an active intrusion. The practical exposure is also uneven: the denial-of-service and VM disk corruption effect works against a default Docker Desktop installation on macOS, while the escalation to arbitrary code execution on the host was demonstrated under non-default configurations, so organisations that have customised Docker Desktop or enabled Docker Extensions carry more risk than those running stock settings.
Attack Surface
Endpoint, Infrastructure, Web Application
Tactics
Execution, Privilege Escalation, Defense Evasion, Impact
Techniques
- T1611 – Escape to Host
- T1068 – Exploitation for Privilege Escalation
- T1059 – Command and Scripting Interpreter
- T1565.001 – Stored Data Manipulation
- T1499 – Endpoint Denial of Service
SuperPRO's Threat Countermeasures Procedures
- Move Docker Desktop on macOS and Windows hosts to the release that carries the vendor fix for CVE-2026-28400 / ZDI-26-150; DMR is compiled into the Docker Backend (com.docker.backend / com.docker.backend.exe), so there is no separately patchable component and the full Docker Desktop update is required.
- Switch off Docker Model Runner in Docker Desktop settings on any workstation that does not need local inference – this removes the unauthenticated listener exposed to containers at model-runner.docker.internal and eliminates the /engines/_configure attack path entirely.
- Prevent untrusted containers from reaching the DMR API by running them on networks created with docker network create –internal, and verify the exposure from a test container with curl -X POST http://model-runner.docker.internal/engines/_configure – a reachable endpoint confirms the host is in scope.
- Disable Docker Extensions in Docker Desktop settings, or restrict installation to Docker Marketplace signed extensions only, because Extensions is the component used to turn the constrained host file write into arbitrary code execution on macOS and to pivot the WSL2 VM-to-host boundary on Windows.
- Alert on HTTP POST requests to the /engines/_configure path on model-runner.docker.internal whose JSON body contains the runtime-flags or raw-runtime-flags properties, since no legitimate workload inside a container should be reconfiguring the host inference runner.
- Monitor for llama-server processes spawned as children of com.docker.backend (macOS) or com.docker.backend.exe (Windows) and flag any command line containing file path arguments that point outside the Docker model cache directory, as this is the observable signature of the argument-injection primitive being used for host file read or write.