Home / Blog / AWS Nitro Sandboxes Put Each EKS Pod in Its Own MicroVM…
Tech News

AWS Nitro Sandboxes Put Each EKS Pod in Its Own MicroVM with One

AWS launched Nitro Sandboxes across all commercial regions, running each opted-in EKS pod inside its own Firecracker microVM at no additional charge.

By Dillip Chowdary • Nov 16, 2026 • Source: InfoQ

AWS Nitro Sandboxes Put Each EKS Pod in Its Own MicroVM with One
Image: InfoQ

At a glance

Product
AWS Nitro Sandboxes
Pricing
No additional charge
Release date
November 16, 2026
Availability
All commercial AWS regions
Minimum Kubernetes
Version 1.36 or later
Hardware limit
No GPU passthrough; max 5 sandboxes per vCPU

Amazon Web Services launched Nitro Sandboxes on November 16, 2026, introducing microVM isolation for Kubernetes pods, Amazon ECS tasks, and Amazon EC2 workloads across all commercial regions at no extra charge. According to InfoQ's report, the feature carves out lightweight virtual machines from parent instances using the Nitro hypervisor to give each container a dedicated guest kernel. Operators enable this isolation in Amazon EKS clusters by adding a single runtimeClassName configuration to standard pod specifications.

This technical guide examines the underlying Firecracker hypervisor architecture, operational modes, cluster prerequisites, and hardware capacity constraints. Cloud architects and Kubernetes operators running multi-tenant platforms, untrusted AI coding agents, continuous integration workers, or remote development environments can use this analysis to evaluate the trade-offs of microVM container sandboxing.

01Firecracker microVM execution and containerd shims

On the parent host, a containerd-shim-nitro-sandbox-v2 process launches a Firecracker microVM as a first-level guest using hardware-assisted virtualization. An internal agent inside the guest VM executes container processes through a nested containerd instance. The parent instance and guest communicate across virtio and vsock channels without sharing memory mappings or kernel state.

AWS authors Carlos Santana and Trishi Mehta emphasized this separation, noting that each pod gets its own operating system completely separate from the host and neighboring workloads. Standard container images start unchanged within the guest environment. Because kernel isolation occurs below the container layer, engineering teams do not need to alter their container images or tooling.

Cluster provisioning depends on the chosen platform tier. On Amazon EKS Auto Mode clusters, the required runtime classes ship preinstalled and provision sandbox-capable nodes on demand. On standard EKS clusters, administrators must build custom sandbox-enabled Amazon Machine Images for their self-managed node groups or Karpenter provisioners before scheduling workloads.

02Isolated and confidential runtime boundary modes

The runtime offers two modes depending on which side of the boundary is trusted. Isolated mode assumes the parent instance is trustworthy while treating the guest as untrusted, preserving kubectl exec, live container logging, and process inspection. Confidential mode treats the parent host as untrusted and moves enforcement inside the virtual machine.

CapabilityIsolated SandboxConfidential Sandbox
Host Trust ModelParent trusted, guest untrustedParent untrusted, guest trusted
Debugging Accesskubectl exec and logs supportedExec and stdin blocked
Policy EngineStandard cluster RBACCedar policy pins pod spec
Cryptographic AttestationNoneNitro Secure Module PCR checks
Secret DecryptionHost-managed credentialsKMS attestation-gated release

In confidential mode, a Cedar policy derived from the pod specification restricts image digests, arguments, environment variables, mounts, and capabilities. The Nitro Secure Module measures image and policy state into Platform Configuration Registers, allowing AWS Key Management Service policies to release secrets only when measurements match. Image layers can remain encrypted during transfer so parent nodes handle only ciphertext.

Attestation guarantees require external verification to take effect. Without a KMS policy or external verifier inspecting register values, recorded measurements remain unverified. Separately, a network-isolated annotation boots sandboxes without virtual network devices, removing DNS, NTP, IMDS, and Pod Identity access while disabling standard HTTP and TCP health probes.

Advertisement

Tech Pulse Daily

Get tomorrow's pulse first

Join engineers who read Tech Pulse before stand-up. Free, weekday mornings.

03Node prerequisites and operator cluster preparation

AWS Container Hero Sandip Das highlighted that isolated mode protection flows in only one direction. An exploited parent instance can inspect unencrypted network traffic, modify the pod /etc/hosts file, or forge credentials. Operators must implement specific cluster prerequisites across their infrastructure before deploying sandboxed pods.

  1. Upgrade control plane Deploy Kubernetes version 1.36 or higher across the cluster.
  2. Enable Nitro Enclaves Configure managed node group launch templates with Nitro Enclaves enabled so pods do not sit in Pending.
  3. Migrate workload credentials Transition authentication architectures from IAM Roles for Service Accounts to Amazon EKS Pod Identity.
  4. Apply runtime class Set runtimeClassName to isolated-sandbox or confidential-sandbox in pod manifests.

Das advised teams to treat vendor claims regarding near-native execution performance with caution. Production rollouts require independent benchmarking across container startup latency, network I/O throughput, and tail latency profiles. Teams must also verify that their credential pipelines operate reliably under the newly required Pod Identity framework rather than legacy service account bindings.

04Hardware density caps and hypervisor limitations

Hardware constraints represent the primary operational trade-offs of the virtualization tier. Firecracker does not support host device passthrough at launch, preventing GPU-accelerated containers from attaching to either sandbox runtime class. Workloads requiring hardware accelerators must continue running directly within standard unisolated container environments.

Capacity density enforces strict ceilings on parent compute instances. Total allocated sandbox vCPUs cannot exceed five times the physical node vCPU count, and total active sandboxes are similarly capped at five times node vCPUs. These thresholds are cut in half on CPU-overcommit instance families, giving a four-vCPU node capacity for twenty single-vCPU sandboxes or five four-vCPU sandboxes.

Note: Simultaneous multithreading core scheduling prevents two sandboxes from sharing sibling threads on a single physical core, cutting usable capacity on x86 processors roughly in half while AWS Graviton processors remain unaffected due to single-threaded cores. EC2 instances running sandboxes also lose support for snapshots, restores, and live migration.

05Workload use cases and cloud comparisons

Amazon positions the virtualization tier for untrusted code execution, identifying artificial intelligence coding agents as the primary workload. Additional use cases include multi-tenant platforms, regulated applications, continuous integration workers, and remote development environments. To reduce microVM cold starts, AWS recommends the open-source agent-sandbox framework, which maintains pre-booted warm instance pools capable of provisioning in approximately two seconds.

The architecture provides an alternative to Google Cloud GKE Sandbox options available since 2019. Google offers a gVisor runtime class that intercepts system calls in under 200 milliseconds with selective CUDA GPU support, alongside a Cloud Hypervisor microVM option requiring 250 mCPU and 130 MiB per pod. AWS instead commits fully to hardware microVM isolation, expressing overhead through density caps rather than static per-pod resource allocations.

Frequently asked questions

How do you enable AWS Nitro Sandboxes on Amazon EKS?

Operators set the runtimeClassName field in the pod specification to either isolated-sandbox or confidential-sandbox. Clusters running Amazon EKS Auto Mode provision compatible nodes on demand, while standard clusters require sandbox-enabled custom AMIs.

What is the difference between isolated and confidential modes?

Isolated mode trusts the host instance and permits kubectl exec and logging while isolating guest processes. Confidential mode distrusts the host, enforces pod specifications using Cedar policies, disables exec and stdin, and gates secrets behind Nitro Secure Module attestation measurements.

Can GPU workloads run inside AWS Nitro Sandboxes?

GPU workloads cannot use Nitro Sandboxes because Firecracker does not support host device passthrough. Any container requiring graphics processing units or hardware accelerators must run without the sandbox runtime class.

What are the capacity limits for Nitro Sandboxes on a host?

A host instance can run at most five sandboxes per physical vCPU, and total allocated sandbox vCPUs cannot exceed five times the node vCPU count. These density limits are halved on instance families that use CPU overcommit.

Sources

Dillip Chowdary

Author

Dillip Chowdary

Writes Tech Bytes coverage of AI, engineering, and the tools that actually ship. Editor of Tech Pulse Daily.

Related on Tech Bytes

Advertisement

5-min tech signal

Weekday briefing for engineers who skip the noise.

No spam · Unsubscribe anytime

Advertisement

✈️ CareerPilot

Your AI job-search copilot

Match your resume against live Ashby, Greenhouse & Lever openings — fit scores, job-specific resume optimization and email alerts.

Find matching jobs →

Free Tools

Browse all tools →