Description

Docker CopyEscape, identified as CVE-2026-17106, is a security flaw that can abuse the `docker cp` file-transfer functionality to write files outside the intended container destination. The vulnerability has a CVSS score of 7.1 and may enable code execution on the host system where the Docker CLI operates. Docker has also confirmed that a related issue affects `sbx cp`, the copy-out functionality used by Docker Sandboxes, including environments designed for AI-agent workloads. Since `docker cp` is widely used for retrieving logs, test outputs, build artifacts, and investigation data, exploitation can turn a routine administrative operation into a potential host compromise vector. The attack relies on two weaknesses in Docker’s archive handling process. A container controlled by an attacker can modify its filesystem during the archive-generation process, including replacing an expected directory with a symbolic link after Docker has inspected it. Docker’s container-level locking does not prevent processes inside the container from making such changes, creating a race condition. During extraction, the Docker CLI can then perform security validation against one path while ultimately writing through a symbolic link to another location. This mismatch can allow archive contents to escape the expected destination and overwrite files on the host. The consequences depend on the privileges of the account performing the copy, with privileged automation potentially exposing sensitive or system-owned files. Organizations should upgrade Docker Engine and CLI to 29.7.2 or later, Docker Desktop to 4.86.0 or later, and Docker Sandboxes to 0.38.0 or later. Until updates are deployed, avoid copying files from untrusted containers while they are running, and minimize privileged commands such as `sudo docker cp`. Automated workflows should also be reviewed for unnecessary root-level permissions. If a container is suspected of malicious activity, conduct evidence collection from an isolated or disposable environment rather than a production workstation. Because technical research and proof-of-concept information is publicly available, affected environments should be patched promptly. At present, there is no confirmed evidence of widespread exploitation of this vulnerability in the wild.