Researchers at Trend Micro's Zero Day Initiative published a full exploitation chain targeting Docker Model Runner, revealing how unauthenticated argument injection can lead to container escapes and arbitrary code execution on host systems.
The Argument-Injection Primitive
According to the Zero Day Initiative, Docker Model Runner features an unauthenticated REST API reachable from any container at model-runner.docker.internal.
The endpoint accepts JSON properties whose values are appended as command-line arguments for llama-server, allowing an unprivileged container to supply arbitrary arguments to the inference process running on the host.
Escaping to the Host via Docker Extensions
Researchers detailed how file-operation flags in llama-server can be leveraged to read and write host files. On macOS, a Seatbelt profile restricts operations, while Windows Job Objects impose no filesystem restrictions.
By combining log-file writes with metadata pollution in GGUF models, attackers can overwrite Docker Extension files and achieve arbitrary code execution on the host when a user opens the extension UI.
Patches and Vendor Response
Docker addressed CVE-2026-28400 in DMR v1.0.16 and Docker Desktop version 4.61.0 through validation and allowlists for backend flags.
A subsequent macOS Seatbelt sandbox bypass, tracked as ZDI-26-451, was mitigated after researchers reported it in February 2026 by denying Mach service lookups for launch services.
What to do
- Update to Docker Desktop version 4.61.0 or later, which includes DMR v1.0.16 and the fixes for CVE-2026-28400.
Key facts and where they come from
- CVE-2026-28400 is an unauthenticated argument-injection vulnerability in Docker Model Runner.
In November 2025, we reported ZDI-26-150 / CVE-2026-28400, a denial-of-service (on macOS) stemming from an unauthenticated argument-injection vulnerability in DMR.
- On macOS and Windows, the DMR inference process runs on the host rather than inside the VM.
on macOS and Windows the inference process does not run inside the VM; it runs on the host, as a child of the Docker Backend
The catch
- The announcement does not say whether the underlying vulnerabilities were actively exploited in the wild prior to public disclosure.
- The Windows exploitation chain requires obtaining the host username, which researchers showed can be leaked via UNC paths or error logs.
Reader poll: Should local AI inference tools run directly on the host or inside restricted virtual machines by default?
A show of hands from readers, not a scientific poll. Say why in the comments.
