Docker Launches MicroVM Sandboxes for AI Agents Across 3 Major Platforms

7 articles · Updated · InfoWorld · May 27

Docker’s new Sandboxes give each container its own microVM, Docker daemon and kernel, aiming to let AI agents build, run and modify software without exposing the host system.
The design targets a gap between regular containers and full virtual machines: containers are lighter but less isolated, while VMs are safer but too heavy for short-lived agentic workloads.
Runtime-scoped access to directories, network endpoints and secrets lets agents use standard Docker commands in isolation, though they can still consume bandwidth and compute by pulling or building excessively.
The same microVM approach also fits malware analysis, untrusted third-party code execution and isolated build pipelines, extending the product beyond its initial AI-agent focus.
Docker says the cross-platform system runs natively on Linux KVM, macOS Hypervisor.framework and Windows Hypervisor Platform, with custom sandbox templates and experimental kits available.

With microVM tech existing for years, what makes Docker's new Sandboxes the breakthrough moment for securing AI agents?

How does Docker's solution ensure its 'full isolation' can truly contain a rogue AI agent in the long run?

Beyond just AI, will the 'container-in-a-VM' model become the new standard for running all untrusted code?