Docker has quietly transformed from “the container company” into the security and runtime layer for agentic AI development. If you’ve checked docs.docker.com lately, you’ve probably noticed the homepage no longer leads with Swarm or Compose tutorials, it leads with questions like “Can I run my AI agent in a sandbox?” and “What are Docker Hardened Images?” Below is a breakdown of the newest Docker features, tools, and documentation updates worth knowing about right now, with hands-on CLI examples for each.

1. Docker Sandboxes: Isolated MicroVMs for AI Coding Agents

Docker Sandboxes give every AI coding agent its own disposable microVM, complete with a private filesystem, its own Docker daemon, and an isolated network stack, accessed through a new sbx CLI.

Install and sign in:

# macOS install
brew trust docker/tap
brew install docker/tap/sbx
sbx login

On first login you’ll choose a default network policy (Open, Balanced, or Locked Down). Then launch an agent in a sandbox from your project directory:

cd ~/my-project
sbx run --name my-sandbox claude

Check what’s running and inspect network rules:

sbx ls
sbx policy ls
sbx policy allow network registry.npmjs.org

Store a credential (like a GitHub token) so the agent can open pull requests without you pasting secrets into the session:

sbx secret set -g github -t "$(gh auth token)"

Clean up when you’re done:

sbx stop my-sandbox
sbx rm my-sandbox

2. Docker Hardened Images (DHI): Near-Zero-CVE Base Images

DHI gives you minimal, distroless-style images with signed SBOMs and SLSA Level 3 provenance. You can pull and run one exactly like a normal image:

docker login dhi.io
docker pull dhi.io/python:3.13
docker run --rm dhi.io/python:3.13 python -c "print('Hello from DHI')"

The really compelling part is comparing a hardened image against its standard counterpart using Docker Scout:

docker scout compare dhi.io/python:3.13 \
  --to python:3.13 \
  --platform linux/amd64 \
  --ignore-unchanged

In Docker’s own example, this comparison showed the hardened Python image dropping from 412MB to 35MB and cutting known CVEs to zero, while removing over 500 unnecessary packages.

3. Gordon: Docker Desktop’s AI Agent, Now GA

Gordon is Docker Desktop’s built-in AI assistant, now generally available with persistent memory and live-streamed command output. You can invoke it straight from the terminal:

docker ai "why is my container exiting immediately?"

Gordon can also be launched in an interactive TUI session for longer troubleshooting conversations, and it now surfaces contextual hints automatically whenever a docker build, docker run, or docker compose up command fails.

4. MCP Toolkit and MCP Gateway: Centralized Control for AI Tool Access

Instead of configuring the same MCP server separately for Claude, Cursor, and every other AI client, Docker’s MCP Gateway runs servers as isolated containers and routes requests through a single point of control. If you’re on Docker Engine without Desktop, install the gateway plugin manually:

chmod +x ~/.docker/cli-plugins/docker-mcp
docker mcp --help

The gateway is started with a specific profile, which determines which of the 300+ catalog servers are available to your clients:

docker mcp gateway run --profile web-dev

5. Docker Model Runner: Local LLMs the Docker Way

Model Runner lets you pull, run, and serve LLMs locally using OpenAI- and Ollama-compatible APIs. On Docker Engine, install the plugin and run your first model:

sudo apt-get update
sudo apt-get install docker-model-plugin
docker model version
docker model run ai/smollm2

Tune context size for a specific model:

docker model configure --context-size 8192 ai/qwen2.5-coder

Publish your own fine-tuned or custom model as an OCI artifact:

docker model tag ai/smollm2 myorg/smollm2
docker model push myorg/smollm2

# Or package a raw GGUF file and push it directly
docker model package --gguf "$(pwd)/model.gguf" --push myorg/mistral-7b-v0.1:Q4_K_M

6. docker pass: Secrets Management from the CLI

Docker’s newer docker pass CLI manages secrets and injects them into host commands rather than leaving them sitting in shell history or .env files:

docker pass set OPENAI_API_KEY --force
docker pass ls
docker pass run -- printenv OPENAI_API_KEY

7. A Steady Cadence of Security Patches

Anyone running Docker Desktop in production environments should keep an eye on the CVE fixes rolling out almost monthly, covering issues in the grpcfuse kernel module, Model Runner’s inference backends, and OCI registry client vulnerabilities. Check current status any time with:

docker version
docker scout cves your-image:tag

Final Thoughts: Docker in 2026 Is an AI Security Platform First

Between Docker Sandboxes, Docker Hardened Images, Gordon, and the MCP Toolkit, Docker is positioning itself as the governance and isolation layer that lets developers safely adopt autonomous coding agents without handing them the keys to the host machine. For teams evaluating their container strategy in 2026, these are the features worth testing first, and every command above is copy-paste ready to try today.

For the full, up-to-date changelog and CLI reference, always check the official Docker Desktop release notes and Docker CLI docs.

References and Further Reading

All the CLI commands and features above were verified directly against Docker’s official documentation. For deeper dives into any of these topics, the following pages are worth bookmarking:

• Docker Desktop Release Notes — the full, versioned changelog for every Docker Desktop release, including Sandboxes, Hardened Images, and Model Runner updates.
• Docker Sandboxes: Get Started — install steps, the sbx CLI reference, and policy configuration for isolated AI agent microVMs.
• Docker Hardened Images (DHI) Overview — what DHI is, why it exists, and how it reduces CVEs compared to standard base images.
• DHI Quickstart — a hands-on walkthrough for pulling, running, and migrating to a hardened image.
• Docker Model Runner — run and serve local LLMs through Docker Desktop with an OpenAI-compatible API.
• Model Runner: Get Started — step-by-step setup, model pulling, and packaging custom GGUF models.
• MCP Catalog and Toolkit — Docker’s curated catalog of Model Context Protocol servers.
• MCP Gateway — securely proxy, log, and manage MCP server connections for AI agents.
• Docker Blog — announcements, tutorials, and roadmap context straight from the Docker team.