AWS Rolls Out Nitro Sandboxes to Give Containers Their Own VM Security Boundary
TL;DR — Key Takeaways
- AWS Nitro Sandboxes bring hardware isolation to individual workloads. Supported Amazon EKS pods, Amazon ECS tasks and other EC2 workloads can each run in a lightweight VM with its own kernel while drawing resources from the same EC2 instance.
- Each sandbox runs alongside its parent EC2 instance. The Nitro hypervisor enforces the boundary directly, and AWS says the design delivers near-native compute and memory performance.
- EKS and ECS offer two sandbox modes. Isolated mode protects the host and neighboring workloads from a compromised sandbox. Confidential mode also protects workloads from host administrators and supports cryptographic attestation before secrets are released.
- The feature carries no additional charge. Customers still pay for EC2 capacity and any applicable EKS Auto Mode or ECS Managed Instances fees.
We’ve always needed as much security as we could get, but the rise of AI agents has made this more important than ever.
To address AI security concerns, Amazon Web Services (AWS) has introduced Nitro Sandboxes. This workload-isolation feature gives individual Kubernetes pods, Elastic Container Service (ECS) tasks, and other Elastic Compute Cloud (EC2)-hosted workloads their own lightweight virtual machines (VMs) while sharing an EC2 instance.
Ringing any bells? If you’re an AWS user, it should. Nitro Sandboxes extend AWS’s Nitro virtualization architecture, which dates back to 2017, to individual container workloads. So, while Nitro Sandboxes are new, Nitro itself has not quite a decade of production use behind it. In other words, you can trust it.
The Nitro Sandboxes rollout targets multitenant applications, build jobs, and AI agents executing generated code. In other words, all the workloads where a shared operating-system kernel can create an unwelcome security dependency between otherwise separate containers.
The feature brings AWS’s Nitro hypervisor boundary down to individual workloads. Rather than sharing the parent instance’s kernel, a sandbox runs its own kernel inside a hardware-isolated VM. Crucially, that VM runs alongside the parent EC2 instance as a first-level Nitro guest, not as a nested VM inside it.
Nitro Sandboxes is now available in all commercial AWS Regions on supported instances, with no additional charge for the sandbox capability. Customers still pay for EC2 capacity and applicable EKS Auto Mode or ECS Managed Instances fees.
A Smaller Hypervisor, a Harder Boundary
The rollout builds on Nitro’s separation of core virtualization functions from networking, storage, and other infrastructure services.
In a Techstrong TV interview about the underlying Nitro architecture, Mark Ryland, AWS’s director of security, said, “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. But we offload from the main CPUs all of the other software that you need for virtualization.”
That offloading moves substantial infrastructure functionality away from the processors executing customer workloads. As Ryland explained, “We offload the virtual networking, virtual storage, and all the other components run on separate computers co-located in the same physical enclosure, but they’re actually different computers with different code, much more protected from the core compute that customers then run.”
You can use Nitro Sandboxes to apply that architecture to workloads smaller than an EC2 instance. In Sandboxes, a parent instance launches and manages sandboxes and supplies their compute, memory, and I/O resources, while the Nitro hypervisor enforces separation between them. AWS claims the first-level guest design avoids nested virtualization and delivers near-native compute and memory performance.
Ryland also described a defensive design technique intended to limit what an attacker could learn even after finding a way to misuse the hypervisor. “We used a technique that’s called secrets hiding,” he said, explaining that the goal is “to minimize the amount of information that the hypervisor even has access to in the first place.”
Isolated and Confidential Modes
Nitro Sandboxes offers two security models. The difference between them is whether the parent instance is trusted. Isolated mode protects the parent and neighboring workloads from a compromised sandbox. The parent instance and its administrators remain trusted, preserving operational access for workloads that need stronger containment but not protection from the infrastructure operator.
Confidential mode is for workloads where you don’t trust the parent instance, such as when you run jobs on an EC2 instance managed by another team. It adds in-sandbox policy enforcement, restrictions on administrative operations, and cryptographic attestation that can verify the workload before releasing secrets through services such as AWS Key Management Service (KMS).
Confidential mode is for when you’re handling sensitive data or model weights on another team’s infrastructure. But it doesn’t magically secure your data. You must still configure a verifier or appropriate AWS KMS policies to check the attested measurements before putting sensitive material within it.
Nitro Sandboxes: Similar Idea to Kata Containers, Different Infrastructure
Nitro Sandboxes isn’t the first technology to combine container workflows with VM isolation. Credit for that goes to Kata Containers, an Apache 2.0–licensed open-source runtime. It also places container workloads inside lightweight VMs with separate guest kernels. Unlike Nitro Sandboxes, which are tied to AWS virtualization, Kata supports multiple virtualization backends, including QEMU, Cloud Hypervisor, and AWS Firecracker, an AWS-specific take on the Linux kernel’s built-in Kernel-based VM (KVM).
Both approaches use hardware virtualization. The difference lies in how their software works. Kata supplies a configurable, portable runtime, while Nitro Sandboxes provide an AWS-native boundary directly under the Nitro hypervisor. A conventional Kata VM running inside a virtualized Kubernetes worker requires nested virtualization, while deploying Kata on bare metal avoids that nesting. Nitro’s sandbox is instead a sibling of the parent EC2 instance.
For AWS users, Nitro virtualization is baked in, so you don’t need to worry about deploying a separate program. Kata’s advantage is its flexibility. You can choose your infrastructure and virtualization stack rather than depend on AWS’s service-specific implementation.
Other companies integrate Kata into their container and AI offerings. Jessica Forrester, Red Hat’s Chief Architect of AI Catalyst Engineering, for example, explained at KubeCon North America 2026 that Red Hat combines OpenShell with Kata so that “Our agents, all of our agents, get their own kernel. Even then, we don’t actually give our agents real keys. They look real to them, but they open nothing, and they’re almost like a prop. OpenShell is swapping in the real key on the way out of the Kata door. This way, if someone convinces our agent to hand over its keys some other way, all it’s handing over is a prop.”
For AWS users, the practical change is that they can now select a hardware VM boundary at pod or task granularity within familiar orchestration workflows, offering a stronger separation model than ordinary shared-kernel containers for teams running potentially hostile code, without requiring a separate EC2 instance for every workload. For those outside the AWS sphere, Kata container-based offerings, such as those Red Hat is working on, may be more interesting.
Frequently Asked Questions
What are AWS Nitro Sandboxes?
Nitro Sandboxes are lightweight, hardware-isolated VMs for individual workloads. Each sandbox has its own operating system kernel, which reduces the risk that a compromise spreads through a shared kernel.
Who might use Nitro Sandboxes?
They’re designed for workloads that need stronger isolation, such as AI agents executing generated code, software build jobs and applications that run workloads from multiple customers on shared infrastructure.
What’s the difference between isolated and confidential mode?
Isolated mode treats the host and its administrators as trusted. Confidential mode is designed for workloads that also need protection from host administrators. With the appropriate AWS KMS key policy, teams can require specific attestation measurements before a sandbox can decrypt sensitive data.


