Cloud Services Containers as a Service (CaaS)

AWS Nitro Sandboxes Give Individual Workloads a VM-Level Boundary

Amazon Web Services has introduced Nitro Sandboxes, a workload-isolation feature that places supported container and other EC2 workloads inside lightweight virtual machines with their own operating system kernels. The design gives individual Kubernetes pods, Amazon ECS tasks and other EC2-hosted workloads a hardware-enforced boundary while allowing them to use resources from the same EC2 instance.

AWS is positioning the feature for environments where workloads may run untrusted or generated code, including multitenant applications, software build jobs and AI agents. Conventional containers share a host operating-system kernel, which can make a kernel compromise a security concern for neighboring workloads. Nitro Sandboxes instead give each sandbox a separate kernel and rely on the Nitro hypervisor to enforce separation.

Workloads run as Nitro guests

A parent EC2 instance launches and manages its sandboxes and provides their compute, memory and I/O resources. Each sandbox runs alongside the parent instance as a first-level Nitro guest, rather than as a virtual machine nested inside the parent. AWS says this avoids nested virtualization and delivers near-native compute and memory performance.

The approach extends AWS’s Nitro virtualization architecture, introduced in 2017, to workloads smaller than an EC2 instance. Nitro separates core virtualization functions from services such as networking and storage, offloading infrastructure work from the processors running customer workloads. For Sandboxes, the hypervisor supplies the hardware isolation boundary between the parent and individual workload environments.

The feature is available in all commercial AWS Regions on supported instances. AWS says Nitro Sandboxes carries no additional charge. Customers continue to pay for EC2 capacity and any applicable fees for EKS Auto Mode or ECS Managed Instances.

Two modes define the trust boundary

Amazon EKS and Amazon ECS support two sandbox modes. Isolated mode is intended to protect the parent instance and neighboring workloads if a sandbox is compromised. It treats the parent instance and its administrators as trusted, so it does not aim to shield a workload from those administrators.

Confidential mode is designed for workloads that must also be protected from the host environment, such as jobs running on an instance managed by another team. It adds policy enforcement within the sandbox and restrictions on administrative operations. It also supports cryptographic attestation, which can be used to verify a workload before allowing it to access secrets through services such as AWS Key Management Service.

Attestation alone does not release sensitive information safely: teams need a verifier or appropriate AWS KMS policies configured to check the sandbox’s measurements before secrets are made available. That requirement is important for deployments involving sensitive data or model weights, where the host is not considered trustworthy.

Relationship to Kata Containers

Nitro Sandboxes shares a broad model with Kata Containers, an Apache 2.0-licensed open-source runtime that also runs container workloads in lightweight virtual machines with separate guest kernels. The technologies differ in portability and deployment. Kata supports multiple virtualization backends, including QEMU, Cloud Hypervisor and AWS Firecracker, while Nitro Sandboxes is an AWS-native feature built on the Nitro hypervisor.

A Kata virtual machine running inside a virtualized Kubernetes worker generally uses nested virtualization; running Kata on bare metal avoids that arrangement. Nitro Sandboxes instead runs as a sibling guest alongside its parent EC2 instance. AWS customers can use the service within its supported environment, while Kata offers a configurable runtime that can be used across different infrastructure and virtualization stacks.

For AWS users, the feature makes a VM-level boundary available at pod or task granularity without requiring a separate EC2 instance for every workload. It can strengthen containment for code that may be hostile or unpredictable, but mode selection and configuration still determine which parties are trusted and whether sensitive material is protected. Organizations seeking portability beyond AWS may find Kata’s broader backend support more relevant.