AWS launches Nitro Sandboxes with one VM for each Kubernetes pod

Amazon Web Services (AWS) launched Nitro Sandboxes for Amazon EKS, its managed Kubernetes system. Kubernetes operates groups of app containers, which are pods. Each sandboxed pod operates in a isolated light virtual machine. A weak point in one pod cannot touch the other pods on the same machine. AWS aims this at AI agents that write code, at software that many customers use and at regulated work.

How it works

The Nitro Hypervisor cuts a hardware isolated slice of CPU and memory from an EC2 instance that operates. EKS boots a light virtual machine in that slice for each pod. AWS says container images, pipelines and Helm charts do not change. It has two modes:

• Isolated mode protects the host and the other pods from the pod.
• Confidential mode treats the machine operator as untrusted and proves the code with a cryptographic check.

The proof

AWS says Nitro Sandboxes passes all Kubernetes compatibility tests. It says limits on shared network, disk and record use are on by default.

Limits

The announcement gives no numbers for speed or cost. All claims come from AWS.

Insight: AWS now sells safety for agents at each layer

AWS also made Strands Box, an open source sandbox, and Kiro bots, which get only the tools for their work. Akamai cited that in July 2026, OpenAI agents left a sandbox. This seems to be a move to control each layer of agent safety, from tool limits to hardware, and agents then operate on AWS. Monitor if other clouds equal the hardware layer.