AWS Nitro Sandboxes isolate each EKS pod in its own microVM
The runtime class runs opted-in workloads in Firecracker microVMs with separate guest kernels. It is available in all commercial regions at no additional charge, but standard EKS requires a sandbox-enabled AMI.

Amazon Web Services has released Nitro Sandboxes, a virtualization feature that places opted-in Kubernetes pods, ECS tasks, and EC2 workloads inside their own lightweight virtual machines with separate guest kernels, according to InfoQ. In Amazon EKS, opting in takes one pod field: runtimeClassName: isolated-sandbox or confidential-sandbox.
Amazon offers the feature in every commercial region without an extra fee. The company targets environments where the tenant or the code is not trustworthy, with coding agents as the top example. Carlos Santana and Trishi Mehta, writing on an AWS blog, said each pod receives "its own operating system, completely separate from the host and from every neighbor."
According to the EKS documentation, the parent instance runs a process called containerd-shim-nitro-sandbox-v2, which starts a Firecracker microVM as a first-level guest using hardware-assisted virtualization. Inside the VM, an agent runs the pod's containers through its own containerd; the parent and guest communicate over virtio and vsock. The sandbox does not share kernel or memory mappings with the parent instance or any other sandbox.
EKS Auto Mode includes the two RuntimeClasses in every cluster and creates sandbox-capable nodes when needed. Standard EKS requires teams to build a sandbox-enabled AMI for node groups or Karpenter.
Isolation modes
Isolated mode trusts the parent and treats the guest as untrusted. kubectl exec, logs, and process listing continue to work. Confidential mode assumes the parent cannot be trusted. In confidential mode, a policy written in Cedar derives from the pod specification and fixes image digests, command-line arguments, environment, volume mounts, and capabilities; exec and stdin are disabled. The Nitro Secure Module records the image and policy in PCRs, and a KMS key policy releases secrets only to a sandbox whose measurements match. Image layers can be encrypted. The documentation warns that without a KMS policy or an external verifier that checks the PCR values, confidential guarantees are incomplete. A network-isolated annotation boots the sandbox without any network device, removing DNS, NTP, IMDS, and Pod Identity and breaking HTTP and TCP probes.
Operator cautions and limits
Sandip Das, an AWS Container Hero, wrote in a LinkedIn analysis that isolated mode protects in one direction only. A compromised parent instance could still observe or tamper with unencrypted traffic and DNS, write the pod's /etc/hosts, and create Pod Identity credentials for any pod on that node. Das lists prerequisites the AWS blog post leaves out: Kubernetes 1.36 or later, and managed node groups must have Nitro Enclaves enabled in their launch template. Without that setting, sandboxed pods remain Pending. He adds that Pod Identity, not IRSA, is required for credentials, and advises treating AWS's near-native performance description as a vendor claim and measuring startup latency, I/O throughput, and tail latency.
Amazon names multi-tenant SaaS, regulated workloads, CI builds, and remote IDEs among the target uses. The company directs users to the open-source agent-sandbox project, whose warm pools deliver a pre-booted VM in roughly two seconds. GPU and device passthrough is unavailable at launch because Firecracker cannot pass host devices to a guest, leaving GPU pods outside the sandbox class.
Density has two caps: total sandbox vCPUs and the number of sandboxes each must not exceed five times the node's vCPU count, and that limit is halved on CPU-overcommit types. A four-vCPU node can host up to twenty one-vCPU sandboxes or five four-vCPU sandboxes. Simultaneous multithreading reduces usable capacity because core scheduling stops two sandboxes from sharing sibling threads, leaving roughly half the advertised vCPUs available; Graviton, with one thread per core, is unaffected.
For comparison, Google's GKE Sandbox has provided a gVisor RuntimeClass starting in 2019; the company later added a microVM class built on Kata Containers and Cloud Hypervisor. gVisor catches syscalls in a user-space kernel, boots in under 200 milliseconds, and works with CUDA on certain GPUs. Google's microVM option charges a fixed 250 mCPU and 130 MiB for each pod, starts in one to two seconds, and lacks GPU support. AWS chose the VM approach and measures cost by density and threads, not per-pod resources.
On EC2, Amazon Linux and Firecracker receive native support, with patches available for other virtual machine monitors. The EC2 User Guide lists supported instances and says snapshot, restore, and live migration are unavailable.
Sources: InfoQ
Kai Brennan covers cloud providers, outages and vendor pricing for Five Nines. He writes from status pages, vendor announcements and published reporting.
Corrections. Report an error to corrections@thefivenines.com. We list substantive corrections on the corrections page.