Docker's new Sandboxes give AI agents safe YOLO mode in isolated microVMs

Docker has released Docker Sandboxes (sbx), a new local CLI tool that gives AI coding agents a safe, disposable execution environment. Each agent runs inside a dedicated microVM with only your project workspace mounted in, meaning Claude Code, Gemini CLI, Copilot CLI, Codex, OpenCode, and Kiro can operate in full YOLO mode (--dangerously-skip-permissions) without any manual review or approval prompts. The sandbox is isolated from your host's filesystem and network, yet agents can still install packages, modify configs, and launch their own Docker containers inside the sandbox. Sandboxes are designed to be faster than traditional VMs, disposable by default, and can be torn down in a single command.

Installation is straightforward: macOS users run brew trust docker/tap && brew install docker/tap/sbx, while Windows users use winget install Docker.sbx. Docker Sandboxes works independently of Docker Desktop and pairs with Docker AI Governance for centralized controls like network restrictions and filesystem policies across an entire team. Docker positions this as infrastructure-level security that lets organizations give agents autonomy without compromising host safety. Early adopters like NanoClaw creator Gavriel Cohen and Warp engineering lead Ben Navetta endorse the approach, with Warp planning to integrate Sandboxes for consistent local and cloud agent environments.