deniz.in

Markets

Weather

Loading weather

· via dev.to (home feed)

AWS Nitro Sandboxes puts each EKS pod and ECS task in its own Firecracker microVM

Launched November 9, 2026, AWS Nitro Sandboxes runs each opted-in EKS pod, ECS task or EC2 workload in its own Firecracker microVM, with coding agents as the first target use case.

AWS Nitro Sandboxes puts each EKS pod and ECS task in its own Firecracker microVM

What launched

AWS introduced Nitro Sandboxes on November 9, 2026, with InfoQ publishing a technical breakdown a week later, on November 16. The feature places opted-in workloads — an EKS pod, an ECS task or an EC2 workload — into a dedicated Firecracker microVM with its own guest kernel, provided by the Nitro hypervisor from the host instance. According to AWS's what's-new announcement, it works on most Nitro-based EC2 instances in all commercial regions and carries no additional charge.

AWS names coding agents first and positions the sandbox as a safe place to execute code generated by LLMs. Other listed scenarios are multi-tenant SaaS platforms, regulated workloads, CI builds and remote IDEs.

How it works

On EKS, opting a pod in takes one field in the pod spec: runtimeClassName set to isolated-sandbox or confidential-sandbox. According to the EKS documentation, the parent node then runs containerd-shim-nitro-sandbox-v2, which starts a Firecracker microVM using hardware-assisted virtualization. An agent inside that VM runs the container through its own containerd over virtio and vsock.

Sandboxes share neither a kernel nor memory mappings with the parent or with each other. Containers start unmodified, and AWS says Pod Identity, the VPC CNI, storage and CloudWatch integration all work without extra configuration.

Two modes, two trust models

Isolated mode assumes the parent node is trustworthy, so everyday operator tooling keeps working: kubectl exec, log access and process listing all behave as usual.

Confidential mode treats the parent as untrusted. A Cedar policy attached to the pod spec pins image digests, arguments, environment variables, mounts and capabilities, and forbids exec and stdin. The Nitro Secure Module measures the image and the policy into PCRs, and per the AWS documentation a KMS key policy can release secrets only to a sandbox whose measurements match. That documentation also describes a gap: without a KMS policy or an external verifier, the measurement is recorded but never verified, leaving confidential mode's guarantees incomplete on their own.

Caveats and limits

InfoQ cites AWS Container Hero Sandip Das on the one-directional nature of isolated mode: a compromised parent can still see unencrypted traffic and DNS, rewrite /etc/hosts and create Pod Identity credentials for any pod on the node.

Das also lists prerequisites: Kubernetes 1.36 or newer, Nitro Enclaves enabled in the launch template of managed node groups — otherwise sandboxed pods never leave the Pending state — and Pod Identity instead of IRSA. He treats AWS's near-native performance claim as unproven and advises measuring startup latency, I/O throughput and tail latency yourself.

Hard limits apply as well. Firecracker cannot pass host devices into the guest, so GPU pods run without the sandbox class at launch. Total sandbox vCPUs and sandbox count max out at five times the node's vCPU count: a 4 vCPU node fits twenty 1-vCPU sandboxes or five 4-vCPU ones, and that budget is cut in half on CPU-overcommit instance types. On machines with SMT, core scheduling prevents two sandboxes from sharing sibling threads of one core, leaving roughly half the vCPUs; Graviton is unaffected.

For context on startup cost, the kubernetes-sigs agent-sandbox project delivers a pre-warmed virtual machine from a warm pool in roughly two seconds, and GKE has offered a gVisor sandbox class since 2019 with startup under 200 ms.

Why it matters

Running LLM-generated code has become a mainstream workload, and the blast radius of a container escape grows with it. Nitro Sandboxes turns hardware-level isolation into a per-pod opt-in — one field in a pod spec, unmodified containers, no extra charge — rather than a separate infrastructure decision, which significantly lowers the adoption barrier.

The fine print matters, though. Isolated mode protects the node from the workload, not the workload from the node, and confidential mode only delivers its stronger guarantees when paired with a KMS key policy or external verification. Teams evaluating it should plan for the version and launch-template prerequisites, the density caps and GPU exclusion, and run their own benchmarks. AWS's offering also arrives into a space where GKE has shipped gVisor-based sandboxing since 2019; the differentiator here is full hardware virtualization per pod via Firecracker.

  • #aws
  • #firecracker
  • #kubernetes
  • #eks
  • #microvm
  • #container-security

Related posts