What's New in Docker Sandbox v0.53.0? Docker released Sandboxes v0.53.0, a tool that runs AI coding agents inside an isolated microVM with its own Linux kernel, a private Docker daemon and a network proxy that enforces an allowlist, according to a hands-on review by Ajeet Raina. In testing, the sandbox's proxy returned 200 for github.com and 403 for an unlisted host under the default "balanced" policy, and the first launch pulled its kit image in about 26 seconds on an 18 CPU / 18 GiB allocation. Raina flagged the mounted workspace as the main residual risk, since an agent can modify files such as .git/hooks and CI config that do not appear in git diff. What's New in Docker Sandbox v0.53.0? Giving Your AI Agents a Safe Place to Play AI coding agents are great right up until they aren't. "YOLO mode" letting the agent run commands without asking permission is where the real productivity lives, but it's also where a single bad rm -rf or a sketchy curl | sh can ruin your afternoon or your laptop . Docker Sandboxes sbx solves this the Docker way: it drops the agent into an isolated microVM with its own kernel, its own Docker daemon, and a network proxy that only lets through the traffic you allow. I spent some time putting the brand-new v0.53.0 https://docs.docker.com/ai/sandboxes/release-notes/?ref=ajeetraina.com 0530 release through its paces. Here's what I found. Installing On macOS it's a one-liner, and notably, Docker Desktop is not required : brew install docker/tap/sbx bash $ sbx version sbx version v0.53.0 2c62838... Before running anything, sbx diagnose is a great sanity check. It walks the whole stack: CLI binary, daemon health, hardware virtualization, filesystem tooling, auth, and network policy: INSTALLATION ✓ CLI binary — found ✓ Daemon — healthy v0.53.0 PLATFORM ✓ Virtualization — supported kern.hv support is 1 CONNECTION ✓ Authentication — authenticated ✓ Global network policy — initialized 17 passed · 3 skipped Spinning up a sandbox I created a throwaway project and launched a bare shell sandbox into it: mkdir -p /tmp/sbx-053-test && cd /tmp/sbx-053-test git init && echo 'print "hello from a microVM" ' app.py sbx create shell --name blog-test /tmp/sbx-053-test First launch pulls the kit image; mine was ready in about 26 seconds , and subsequent runs are cached. The CLI tells you exactly what you're getting: the image digest, the mounted workspace, and the resources allocated 18 CPU / 18 GiB on my machine . The four isolation layers, verified by hand This is the part that matters, so I tested each boundary rather than taking the docs' word for it. 1. A separate kernel bash $ sbx exec blog-test uname -a Linux blog-test 7.0.14 1 SMP PREEMPT ... aarch64 GNU/Linux That's a full Linux guest kernel running on my macOS host via the hypervisor, not a shared namespace. The agent runs as an unprivileged agent user uid 1000 but has sudo inside the box by design . The security boundary is the VM, not in-guest permissions. 2. The workspace is mounted, and writes flow both ways bash $ sbx exec blog-test sh -c 'ls /tmp/sbx-053-test && cat app.py' README.md app.py print "hello from a microVM" Write a file inside the sandbox and the host sees it instantly: bash $ sbx exec blog-test sh -c 'echo "written inside the microVM" from-sandbox.txt' $ cat /tmp/sbx-053-test/from-sandbox.txt written inside the microVM ⚠️ The workspace is the main residual risk. The agent can touch anything in the mounted directory, including .git/hooks/ , CI config, and build scripts, which don't show up in git diff . Worth a post-session glance. 3. A private Docker daemon inside the sandbox This one is my favorite. The sandbox ships its own Docker Engine, fully isolated from the host's: bash $ sbx exec blog-test docker version --format '{{.Server.Version}}' 29.9.3 $ sbx exec blog-test docker run --rm hello-world Status: Downloaded newer image for hello-world:latest Hello from Docker Your agent can build and run test containers all day long without ever touching your host daemon. 4. A network proxy that actually enforces policy With the default balanced policy, I hit an allowed host and a random unlisted one: bash $ curl -s -o /dev/null -w "%{http code}\n" https://github.com 200 $ curl -s -o /dev/null -w "%{http code}\n" https://www.random-unlisted-site-xyz.com 403 And sbx policy log shows exactly why. Every decision is auditable: BLOCKED REQUESTS www.random-unlisted-site-xyz.com:443 No matching allow rule default deny ALLOWED REQUESTS registry-1.docker.io:443 github.com:443 Notice that registry-1.docker.io appears in the allow list. That's the hello-world pull from the test above, routed through the proxy. Nothing reaches the network without going through the policy engine. What's new in v0.53.0 A few highlights stood out beyond the core isolation story. sbx df : local disk accounting. A new command that finally answers "where did all my disk go?" with a docker system df -style breakdown and a reclaimable estimate: bash $ sbx df TYPE TOTAL ACTIVE SIZE RECLAIMABLE Images 113 19 50.29GB 17.43GB 35% Containers 46 1 12.93GB 12.93GB 100% Local Volumes 162 162 4.733GB 0B 0% Claude in Chrome access. Enable the claude.chrome setting and sandboxed Claude Code can drive Claude in Chrome using your subscription OAuth login start it with --chrome . API-key auth alone won't cut it. This is subscription-gated. Automatic loopback port publication for kits. Kits granted ports.publish with approval: none can grab an auto-assigned loopback port without a prompt fixed ports still require approval . Resilience after network changes. Stalled outbound connections now close after ~30s so clients can reconnect cleanly. Handy when you connect a VPN mid-session. There's also a specific fix for TLS-inspecting proxies Zscaler, Palo Alto, F5 that were rejecting ML-DSA certificate handshakes. A couple of breaking changes worth noting if you're upgrading: - The built-in gemini agent name is retired . Existing Gemini sandboxes keep running but can't be sbx move 'd. - sbx run