AWS Launches Nitro Sandboxes, Giving EKS Pods, ECS Tasks and EC2 Workloads Their Own Hardware-Isolated VM
AWS Nitro Sandboxes puts each EKS pod, ECS task or EC2 workload in its own hardware-isolated microVM on a shared instance, aimed at multi-tenant, regulated and AI-agent workloads.
Editor's Note ·
- Correction:
- The article's "What We Don't Know" section states that "Neither AWS's announcement nor its Compute Blog post specifies which Nitro-based instance families, beyond 'most,' support Nitro Sandboxes." This is incorrect. The AWS Compute Blog post does specify the supported instance scope: "Every EC2 instance type that supports Nitro Enclaves, across general purpose, compute optimized, memory optimized, storage optimized, and accelerated families, on AWS Graviton, Intel, and AMD processors."
Overview
Amazon Web Services has introduced Nitro Sandboxes, a feature that gives individual Amazon EKS pods, Amazon ECS tasks and other EC2-hosted workloads their own hardware-isolated virtual machine while still sharing a single EC2 instance, according to AWS’s announcement. The feature is available today on most Nitro-based EC2 instances, in all commercial regions, at no additional charge beyond standard EC2 and applicable EKS Auto Mode or ECS Managed Instances fees, according to AWS and independently confirmed by Cloud Native Now.
What We Know
Nitro Sandboxes run each opted-in workload inside its own lightweight VM managed by the Nitro Hypervisor, giving it “the isolation boundary of a Nitro isolated virtual machine” so that untrusted, multi-tenant or regulated workloads can run safely within a single instance’s resources, according to AWS. Enabling the feature on EKS or ECS is a one-line change, and workloads keep native compatibility with Pod Identity, VPC CNI, EBS/EFS/S3 storage and CloudWatch, AWS says. Customers who want direct control can instead run Nitro Sandboxes on EC2 with their own AMI and virtual machine monitor; Amazon Linux and Firecracker are natively supported, with patches available for other options, per the same announcement.
The feature extends the Nitro System, whose architecture dates back to 2017, down to individual container workloads, as reported by Cloud Native Now. A sandbox runs its own kernel inside a hardware-isolated VM that operates as a first-level Nitro guest alongside the parent EC2 instance rather than as a nested VM, which AWS says allows near-native compute and memory performance, according to both AWS and Cloud Native Now. On the AWS Compute Blog, Alex Graf, a principal engineer on the EC2 team, frames the launch as a response to customers who wanted the isolation of Nitro Enclaves extended to containers: “Operators of a shared cluster have access to all the tasks running in it. A weakness in the shared kernel reachable by one container also exposes every other container on the host,” he writes, adding that “with the current rate of Linux Common Vulnerabilities and Exposures (CVEs), containerized workloads need an architecture that isolates each workload from other workloads and from cluster operators.”
The rollout targets multitenant applications, build jobs and AI agents executing generated code, according to Cloud Native Now; AWS’s own announcement likewise points to running “LLM-generated code in an isolated environment of its own” while sharing a single instance across multiple agentic workloads.
Nitro Sandboxes offers two modes. Isolated mode protects the parent instance and neighboring workloads from a compromised sandbox while keeping the parent instance and its administrators trusted, according to Cloud Native Now. Confidential mode is for cases where the parent instance itself is not trusted — for example, running a job on an instance managed by another team — and adds in-sandbox policy enforcement, restrictions on administrative operations, and cryptographic attestation that can verify a workload before releasing secrets through AWS Key Management Service, per the same report. Cloud Native Now notes that confidential mode “doesn’t magically secure your data”: customers must still configure a verifier or appropriate AWS KMS policies to check attested measurements before placing sensitive material inside a sandbox.
Speaking about the underlying Nitro architecture in a Techstrong TV interview, as reported by Cloud Native Now, AWS director of security Mark Ryland described the design philosophy: “Nitro is a complete re-imagining of how virtualization should happen on an x86 or ARM processor. We strip the hypervisor down to the absolute minimum, essentially software that only splits up the CPU and memory into the different compartments needed for the virtual machines.” Ryland also described a technique he called “secrets hiding,” intended, he said, “to minimize the amount of information that the hypervisor even has access to in the first place.”
Cloud Native Now also places Nitro Sandboxes alongside Kata Containers, an existing Apache 2.0-licensed open-source runtime that likewise places container workloads inside lightweight VMs with separate guest kernels but, unlike Nitro Sandboxes, supports multiple virtualization backends including QEMU, Cloud Hypervisor and AWS’s own Firecracker. At KubeCon North America 2026, Red Hat’s chief architect of AI catalyst engineering, Jessica Forrester, described how Red Hat combines its OpenShell project with Kata so that “our agents, all of our agents, get their own kernel,” issuing them cryptographic keys that “look real to them, but they open nothing, and they’re almost like a prop,” according to Cloud Native Now.
What We Don’t Know
Neither AWS’s announcement nor its Compute Blog post specifies which Nitro-based instance families, beyond “most,” support Nitro Sandboxes, or whether any regions are excluded from the stated worldwide commercial-region availability. AWS has not published independent benchmark figures quantifying the “near-native” performance claim.
Context
Nitro Sandboxes is AWS’s entry into a broader push toward hardware- or kernel-level isolation for AI-agent code execution. The Machine Herald has previously reported on Docker and Deno shipping their own competing AI code sandboxes, and has separately covered Google’s move to donate its gVisor-based agent sandboxing work to the CNCF.