Docker introduces microVM sandboxes for AI coding agents

Docker has released Docker Sandboxes, a new feature designed to provide disposable, isolated microVM environments for AI coding agents. The service supports integration with tools including Claude Code, Gemini CLI, Copilot CLI, Codex, Kiro, and OpenCode, allowing these agents to install packages and run services within a secure containerised environment.

The primary function of the sandboxes is to protect the host filesystem and network by isolating agents within dedicated microVMs. This architecture offers stronger isolation than standard containers but with lower resource overhead than traditional virtual machines. Agents operating within these sandboxes can also spin up additional Docker containers, enabling more complex workflows without compromising the host system.

Docker has also introduced Docker AI Governance, a service that allows organisations to enforce centralised network and filesystem controls across entire teams. This tool enables the definition of policies, including network policies, filesystem rules, and MCP governance, which can be enforced on every developer’s machine to maintain security standards.

The platform includes a feature described as "YOLO mode," activated via the `--dangerously-skip-permissions` flag. This mode grants agents autonomy without approval prompts, a capability Docker describes as essential for speed but notes carries risk without proper guardrails. The service is designed to let agents work unattended while maintaining isolation.

Out of the box, the solution supports the aforementioned major coding agents. Docker positions the release as a response to the need for secure execution environments as AI tools become more prevalent in software development workflows, aiming to prevent accidental or malicious damage to host systems during package installation and service execution.