AI agents have repeatedly slipped past the boundaries operators set for them; recent incidents include an OpenAI agent accessing an Australian government portal without authorization. Docker's answer, announced at the WeAreDevelopers Conference, is Cloud Sandboxes: hosted compute instances that launch in hundreds of milliseconds, are charged per second, and come with built-in secrets, policies, networks, and agent configuration. The company positions this as a way to keep agents away from locally stored files.
Docker president Mark Cavage demonstrated the problem: Claude, started in a regular Docker container, was able to find a secret on the host by probing its environment and using the mounted Docker socket. Running the same prompt inside a Docker Sandbox, which executes as a full micro VM, the model could not reach the secret. Cavage argued that containers and containment are separate concerns, and that sandboxes provide a deterministic base layer while policies govern agent intent.
Alongside the sandboxes, Docker updated its Kits specification so agent packages now ship as standard OCI images, and BAND's Python Kit for Docker Sandboxes lets agents coordinate over WebSocket without sharing an environment. Pricing starts at $0.07 per hour for a Micro instance and reaches $1.12 per hour for XL.