The Pulse
Researchers detail Docker Model Runner container-escape paths
Zero Day Initiative researchers found that an argument-injection flaw in Docker Model Runner could lead to host code execution under non-default configurations. Their analysis also identified a separate macOS sandbox escape for which Docker issued an update.

Researchers found that a container with default privileges could reach files on a Docker Desktop host through Docker Model Runner and, under non-default configurations, use that access to execute code on the host. The Zero Day Initiative's Nitesh Surana detailed the attack paths in research published on November 18.
The finding builds on an argument-injection flaw tracked as CVE-2026-28400, which the researchers had reported as a denial-of-service issue on macOS. Further analysis found a route from a constrained file write to host code execution through Docker Extensions, as well as separate escape paths involving the macOS sandbox and the Windows virtual machine boundary.
From container to inference process
Docker Model Runner pulls, runs and serves models locally as part of Docker Desktop. The research covered Linux containers running through Docker Desktop on macOS and Windows; Docker Desktop for Linux and Windows containers on Windows were outside its scope.
On macOS and Windows, the inference process runs on the host rather than inside the Linux virtual machine that houses the containers. It runs with the privileges of the Docker Desktop user and is placed in an operating-system sandbox.
The Model Runner API is unauthenticated and reachable from containers through an internal address. Surana found that a container with default privileges could use its configuration endpoint to supply command-line arguments to the host's inference process.
The configuration is applied when a runner is next started for the relevant model and mode. If a runner is already active, the changed configuration is ignored for that existing process.
File access differs by system
At the time of the research, Model Runner used a llama.cpp inference server. Some of its command-line options read files or write inference logs, making the files available to the process important to the researchers' analysis.
On macOS, the allowed paths included Docker Desktop's data directory. Surana found that directing the log output to the virtual machine's data disk could overwrite it; the disk contains the current Docker Desktop user's containers, images, volumes and image build cache.
On Windows, the Job Object used to constrain the inference process imposed no filesystem restrictions beyond those associated with the user's privileges, according to the research. The process therefore had the Docker Desktop user's filesystem access, subject to the user's access controls.
File paths presented another obstacle on Windows because the working directory was not deterministic. The researchers described ways to obtain the Docker Desktop username from a container without user interaction, including through a network path that disclosed it during a connection attempt.
A separate sandbox issue
Surana also described an escape involving an overly permissive macOS Seatbelt sandbox configuration and another that crossed the Windows WSL2 virtual machine-to-host boundary through Docker Extensions.
A separate July advisory, ZDI-26-451, describes the macOS sandbox issue. It requires an attacker first to be able to run low-privileged code within the sandbox, after which the flaw can be used to execute code on the host as the current user. Docker has issued an update to correct that vulnerability.
Sources
Related stories
- Three High-Severity vLLM Flaws Expose Serving Systems to Remote Attacks
- Firecrawl kit gives Docker sandbox agents controlled web access
- Microsoft researchers show neuron edits can backdoor AI agents
- Aikido finds reachable secrets in 13 of 50 AI security incidents
- GitHub and Microsoft pilot turns AI findings into 99 checked reports