Docker introduces Cloud Sandboxes: a solution that builds on local AI testing environments and is immediately accessible to developers.
Docker launches Cloud Sandboxes. This is a solution that allows AI agents to operate in isolated environments in the cloud, even when the developer’s laptop is turned off. The service builds on the local sandboxes the company introduced earlier this year and is available now.
With Cloud Sandboxes, Docker aims to provide a tool for running AI workflows as they become increasingly time-consuming and are no longer limited to sessions of just a few minutes. Developers can execute tasks that take hours, such as large-scale refactoring or test suites, without requiring their own machine to remain active.
Isolation
Cloud Sandboxes utilize the same microVM isolation as Docker’s local sandboxes. Each agent is assigned its own kernel and Docker daemon, segregated from the rest of the infrastructure, ensuring that files, network access, and secret keys remain protected if an agent fails to comply with the rules.
In addition to Cloud Sandboxes, Docker is also releasing a new version of Kits, the open specification that packages a sandbox with its associated agent, tools, and access rules as a single shareable package. Kits are now built as standard OCI images, without a proprietary format. Docker has announced it will submit the Kits specification to the Cloud Native Computing Foundation.
From laptop to cloud
Cloud Sandboxes are intended for developers who want to scale AI workflows without having to provide their own infrastructure or pay for unused capacity. They can start a sandbox on their laptop and then move it to the cloud with a single command, or start directly in the cloud.
The environments boot within a few hundred milliseconds, with secret keys, policies, MCP gateways, and agent configurations pre-integrated. Computing power scales from one to sixteen vCPUs and is fully managed by Docker. Developers write their commands once and execute them under the same controls, both locally and in the cloud.
AI resists the sandbox
The importance of secure and isolated testing environments for AI is becoming clearer every day. Docker is targeting developers who want to work with AI locally but can no longer or no longer wish to run everything on their laptops.
However, major AI companies also test their models in sandboxes, with varying degrees of success. AI appears to have a tendency to escape restrictive environments in its search for answers. Even some cybercriminal behavior is not foreign to LLMs in such cases. OpenAI has even briefly paused testing of its most advanced models following yet another escape.