AWS Nitro Sandboxes run pods, ECS tasks, and EC2 workloads in their own lightweight VMs with separate kernels, at no extra charge for the feature.

Shared kernels have always been the weak spot in container security. Every pod on a node trusts the same operating system underneath it, so one escape can become a problem for everything nearby. AI agents that execute generated code make that trade-off much harder to accept. Amazon Web Services has responded with Nitro Sandboxes, which give individual workloads their own hardware-isolated virtual machine and kernel.
The feature targets Amazon Elastic Kubernetes Service (EKS) pods, Amazon Elastic Container Service (ECS) tasks, and other Elastic Compute Cloud (EC2) workloads. Each one can run in a lightweight VM while drawing compute, memory, and I/O from the same parent EC2 instance. You get VM-level separation without provisioning a dedicated instance for every job.
Why the Nitro Hypervisor Matters Here
The interesting part isn't the sandbox itself. It's where the sandbox sits. AWS built Nitro sandboxes on the same virtualization architecture that underpins its EC2 fleet, a design that dates to 2017. That gives Nitro roughly eight years of production history, which is a meaningful track record for a boundary that now carries untrusted code.
Nitro sandboxes don't run nested inside the parent instance. Each one is a first-level Nitro guest that sits alongside the parent, so there's no second layer of hypervisor between the sandbox and the hardware. AWS says this avoids the performance penalty of nested virtualization and delivers near-native compute and memory performance. That claim deserves real-world benchmarking, but the architecture at least removes the most obvious overhead problem.
AWS has also been moving functionality out of the hypervisor for years. In a Techstrong TV interview, Mark Ryland, AWS's director of security, described the approach this way: "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."
Ryland said networking and storage run on separate hardware in the same physical enclosure, with different code from the compute that customers run. He also pointed to a technique called secrets hiding, whose goal is "to minimize the amount of information that the hypervisor even has access to in the first place." The idea is to limit what an attacker could learn even after finding a flaw in the hypervisor. Whether that holds up under scrutiny is something only time and independent research will settle, but it's a sensible design principle.
Isolated Versus Confidential Mode
Nitro Sandboxes come in two modes, and the distinction comes down to one question: do you trust the parent instance?
Isolated mode assumes the parent and its administrators are trustworthy. Its job is to keep a compromised sandbox from reaching the host or neighboring workloads. Teams keep normal operational access to the instance, which suits workloads that need tighter containment but don't need protection from the infrastructure operator.
Confidential mode is built for the opposite case. Think of a job running on an EC2 instance that another team manages. In this mode, the sandbox adds in-sandbox policy enforcement, restricts administrative operations, and supports cryptographic attestation. Before a workload receives secrets through services such as AWS Key Management Service (KMS), it has to prove what it is.
Attestation doesn't secure anything on its own, though. Teams still need to configure a verifier or appropriate KMS key policies that check the attested measurements before sensitive material goes into a sandbox. That's real engineering work, and organizations that skip it will get the label without the protection.
How Nitro Compares With Kata Containers
Nitro Sandboxes aren't the first attempt to pair container workflows with VM isolation. Kata Containers, an Apache 2.0-licensed open-source runtime, does the same thing with lightweight VMs and separate guest kernels. The key difference is portability. Kata supports several virtualization backends, including QEMU, Cloud Hypervisor, and AWS Firecracker, so you can pick the stack that fits your infrastructure.
Nitro Sandboxes are narrower by design. They're an AWS-native boundary sitting directly under the Nitro hypervisor, which means AWS users get it without deploying anything extra. The trade-off is lock-in. Kata gives you control over the stack, while Nitro gives you less to manage.
Infrastructure matters too. A Kata VM running inside a virtualized Kubernetes worker needs nested virtualization. Running Kata on bare metal avoids that. Nitro sidesteps the question by making the sandbox a sibling of the parent instance rather than a child of it.
Red Hat is taking the Kata route. Jessica Forrester, Red Hat's Chief Architect of AI Catalyst Engineering, explained at KubeCon North America 2026 that the company pairs OpenShell with Kata. "Our agents, all of our agents, get their own kernel," she said. "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."
What Changes for Teams Right Now
The practical shift is granularity. AWS users can now choose a hardware VM boundary at the pod or task level, inside the orchestration workflows they already run. For teams executing potentially hostile code, that's a meaningful step up from ordinary shared-kernel containers, and it doesn't require a separate EC2 instance for every workload.
Pricing is simple. The sandbox capability carries no additional charge, and it's available in all commercial AWS Regions on supported instances. You still pay for EC2 capacity and any applicable EKS Auto Mode or ECS Managed Instances fees.
My read is that Nitro Sandboxes will matter most to platform teams running multitenant SaaS, CI build farms, and agent platforms, where a shared kernel has always been a quiet liability. The open question is how much of the confidential mode's value depends on teams doing the attestation work properly. Watch for the first independent security evaluations of the sandbox boundary and for early benchmark data on the near-native performance claim. Those will tell you whether this is a niche isolation option or the default for untrusted workloads on AWS.
Discussion
Sign in to join the discussion.