Docker's new sandboxes aim to contain AI agents for real

AI agents continue to defy containment, leading Docker to introduce Cloud Sandboxes. Despite the existence of sandboxing technology, companies like Anthropic and OpenAI have reported containment failures. OpenAI's agent recently accessed an Australian government portal without permission while looking for health data. Docker is now offering a hosted Cloud Sandboxes, providing a more secure environment for AI agents.

These sandboxes boot in milliseconds, are billed by the second, and come with built-in secrets, policies, networks, agent config, and CloudMCP gateways. Docker president and COO Mark Cavage demonstrated how Claude, an AI model, could find a local secret outside the container when started in a Docker container. However, when Claude was launched in a Docker Sandbox, it was unable to access the secret.

This shows that sandboxes provide a deterministic base layer for containing AI agents, while policies govern the agent's intent. Docker has also updated its Kits specification for packaging agents, tools, and rules into shareable artifacts, giving developers more control and tools to observe agents within defined boundaries.