AWS Nitro Sandboxes Give Your Containers Their Own Hardware Walls
AWS Nitro Sandboxes, announced in the November 16 weekly roundup, give EKS pods and ECS tasks their own VM boundary on shared EC2 hosts, at no extra charge. It is the strongest isolation option AWS has offered for container workloads.

AWS Nitro Sandboxes bring hardware isolation to containers
AWS Nitro Sandboxes bring hardware isolation to pods, tasks, and VMs, according to the AWS Weekly Roundup published on November 16. You can give an Amazon EKS pod or Amazon ECS task its own VM boundary while sharing an Amazon EC2 instance. AWS Nitro Sandboxes have no additional charge on supported instances, though EC2 and applicable EKS Auto Mode or ECS Managed Instances fees still apply.
That last sentence is the one that matters. Hardware-level isolation used to mean paying for a dedicated host, which meant paying for capacity you did not use. Nitro Sandboxes give you the isolation boundary without the dedicated host bill. For teams that run sensitive workloads next to untrusted ones, that is a meaningful change.
What Nitro Sandboxes actually do
Containers share a kernel. That sharing is the source of both their efficiency and their risk. A kernel vulnerability in a shared host can, in the worst case, let one container reach another. That is why regulated industries and security-conscious teams have historically run sensitive workloads on dedicated hosts or separate clusters.
Nitro Sandboxes change the trade by putting each pod or task behind its own VM boundary, enforced by the Nitro hypervisor. The workload still shares the physical EC2 instance with other workloads, so you keep the cost efficiency of shared hosts. But each workload gets hardware-enforced isolation from its neighbors, the way a full virtual machine would.
The use cases write themselves. Run a customer-facing service next to a batch job that processes untrusted input. Run workloads from different teams, or different trust levels, on the same cluster. Run a compliance-sensitive workload in a shared environment without giving up the isolation your auditors expect.
The rest of the November 16 roundup
Nitro Sandboxes were the headline, but the weekly roundup carried several other launches worth knowing about. Amazon Cognito now supports OAuth 2.0 token exchange, so you can exchange a user's token for one scoped to a downstream service. Amazon ElastiCache Serverless adds search and aggregations, combining vector, full-text, and hybrid search with server-side aggregations for Valkey.
Amazon OpenSearch Service adds managed cross-Region disaster recovery, letting you pair domains running version 3.7 or later in two Regions within one account. Amazon CloudWatch Logs adds account-level retention policies, so you can set retention across all log groups or groups matching a name prefix. And the next generation AgentCore runtime is now the default, with agent sessions starting in under 2 seconds, scaling up to 8 vCPU and 32 GB, scaling to zero when idle, and billing up to 68 percent less memory.
The AgentCore change connects to the AWS Emulator that AWS launched last week, which lets developers test AWS API interactions locally. AWS is building a full loop for agent development: emulate locally, deploy on AgentCore, and isolate the runtime with Nitro Sandboxes.
When to use Nitro Sandboxes
Not every workload needs hardware isolation. If all your containers run your own code, trust the same team, and handle the same sensitivity of data, the standard container boundary is fine. Adding a VM boundary you do not need is complexity without benefit.
Use Nitro Sandboxes when the trust levels differ. A pod that processes user-uploaded files next to a pod that holds payment data. A third-party integration running in your cluster. A workload that handles regulated data alongside a general-purpose service. Those are the cases where a kernel escape would be catastrophic, and where hardware isolation is worth the small overhead.
Also consider it for multi-tenant platforms. If you run customer workloads in a shared cluster, Nitro Sandboxes give each tenant a hardware boundary without forcing you onto dedicated hosts. That is a strong selling point for SaaS teams that need to show auditors how tenants are separated.
The cost question
AWS says Nitro Sandboxes have no additional charge on supported instances. That is the right price for a feature like this, because charging extra would have pushed teams to skip it. The EC2 and EKS or ECS fees still apply, so the cost is the same as running the workload normally. The isolation is effectively free, which means the only reason not to use it is the operational overhead of configuring it.
Check the supported instance list before you plan a migration. Not every instance type supports Nitro Sandboxes, and the feature is new enough that documentation and tooling are still catching up. Test it in a non-production cluster first, measure the performance overhead on your workloads, and confirm your monitoring and logging still work inside the sandbox boundary.
The bigger picture
AWS is in a race to make the cloud safe for AI agents, and Nitro Sandboxes are part of that race. Agents run code, call tools, and process untrusted input, all of which increase the value of strong isolation. AWS opened its first cloud region in Chile with $4 billion behind it, and the infrastructure buildout is accelerating everywhere. The common thread is that AWS wants every kind of workload, including the risky ones, to run on its platform.
Nitro Sandboxes are a quiet launch, but they solve a real problem. If you run containers with mixed trust levels, try them. The isolation is free, the setup is straightforward, and the alternative is hoping a kernel vulnerability never shows up in your cluster.


