Docker Model Runner Flaws Allow Container Escape

Security researchers say vulnerabilities in the Docker Desktop AI model feature can let containers execute code on the host computer.

Updated on Nov. 18, 2026 in Artificial Intelligence

Docker Model Runner Flaws Allow Container Escape

Security researchers disclosed container-escape vulnerabilities in Docker Model Runner, a Docker Desktop feature for running AI models locally. The researchers say the flaws can lead from a standard container to code execution on the host, and the behavior was reproduced on macOS and Windows under non-default configurations.

Why it matters

Because the feature pulls, runs and serves AI models on a developer's own machine, a container escape could expose the host system to code it was not meant to run. Developers and organizations using Docker Desktop for local AI work have a direct stake in how quickly these issues are addressed.

The disclosure covers a feature that runs AI models locally and was reproduced on macOS and Windows only under non-default configurations. The researchers did not provide figures on how many installations are exposed, and the status of fixes is not stated in the source.

The players

Docker Model Runner

A feature within Docker Desktop that pulls, runs and serves AI models locally on a developer's machine, and the subject of the disclosed vulnerabilities.

The details

Researchers say that from a standard container, the flaws can lead to code execution on the host computer. One of the flaws was previously reported in November 2025 as a denial-of-service issue on macOS, and a separate escape involved an overly permissive macOS sandbox configuration.

Timeline

  1. November 2025: An earlier denial-of-service vulnerability in Docker Model Runner was first reported.

  2. November 18, 2026: Article published describing the container-escape research.

Big Picture

The disclosure extends the earlier November 2025 denial-of-service report on Docker Model Runner into a more serious container-escape finding. It shows that local AI model features built into Docker Desktop are now drawing security scrutiny on both macOS and Windows.

Developers who run Docker Desktop to serve AI models locally could have their host machines exposed if a container escape is exploited. Checking Docker Desktop's update status and limiting non-default configurations are practical steps while the flaws remain under review.

The takeaway

Local AI tools widen the attack surface on developer machines, so keeping container runtimes updated and avoiding permissive sandbox settings matters. Treat locally served models as code that deserves the same isolation as any other untrusted workload.

Further reading

Explore more on Artificial Intelligence and how local AI tools are being built and secured.

Source note: This article includes information reported by IT Security News - cybersecurity, infosecurity news.