CVE-2026-5817 — Docker Desktop vulnerability

Vendor	Product	Version range	Fixed in
docker	docker_desktop	>= 4.62.0 < 4.68.0	4.68.0
docker	docker_desktop	>= 4.62.0 < 4.68.0	4.68.0

GHSA

GHSA-cgqp-ww2v-6rjh: The vllm-metal inference backend in Docker Model Runner on macOS unconditionally sets trust_remote_code=True when loading model tokenizers, and runs w

ghsa_unreviewed·2026-05-26

CVE-2026-5817 [HIGH] CWE-829 GHSA-cgqp-ww2v-6rjh: The vllm-metal inference backend in Docker Model Runner on macOS unconditionally sets trust_remote_code=True when loading model tokenizers, and runs w The vllm-metal inference backend in Docker Model Runner on macOS unconditionally sets trust_remote_code=True when loading model tokenizers, and runs without sandboxing. This causes transformers.AutoTokenizer.from_pretrained() to import and execute arbitrary Python files included in any model pulled from an OCI registry, resulting in arbitrary code execution on the Docker host as the Docker Desktop user when inference is triggered. Any container on the Docker network can trigger this by calling the model-runner.docker.internal API to pull a malicious model and request inference.

VulDB

Docker Desktop up to 4.67.x vllm-metal Inference Backend transformers.AutoTokenizer.from_pretrained inclusion of functionality from untrusted control sphere

vuldb·2026-05-23

CVE-2026-5817 [CRITICAL] Docker Desktop up to 4.67.x vllm-metal Inference Backend transformers.AutoTokenizer.from_pretrained inclusion of functionality from untrusted control sphere A vulnerability classified as critical has been found in Docker Desktop up to 4.67.x. The impacted element is the function transformers.AutoTokenizer.from_pretrained of the component vllm-metal Inference Backend. The manipulation leads to inclusion of functionality from untrusted control sphere. This vulnerability is listed as CVE-2026-5817. The attack must be carried out locally. There is no available exploit. It is recommended to upgrade the affected component.

CVEList

Docker Model Runner container-to-host code execution via unsandboxed trust_remote_code in Python inference backends

cvelistv5·2026-05-22·CVSS 8.8

CVE-2026-5817 [HIGH] CWE-829 Docker Model Runner container-to-host code execution via unsandboxed trust_remote_code in Python inference backends Docker Model Runner container-to-host code execution via unsandboxed trust_remote_code in Python inference backends The vllm-metal inference backend in Docker Model Runner on macOS unconditionally sets trust_remote_code=True when loading model tokenizers, and runs without sandboxing. This causes transformers.AutoTokenizer.from_pretrained() to import and execute arbitrary Python files included in any model pulled from an OCI registry, resulting in arbitrary code execution on the Docker host as the Docker Desktop user when inference is triggered. Any container on the Docker network can trigger this by calling the model-runner.docker.internal API to pull a malicious model and request inference.

CVE-2026-5817: The vllm-metal inference backend in Docker Model Runner on macOS unconditionally sets trust_remote_code=True when loading model tokenizers, and runs without…

Affected

CVSS provenance