๐ November 16, 2026 โข infoq.com
AWS launched Nitro Sandboxes on November 9, 2026, and InfoQ published a technical breakdown on November 16. It runs an opted-in EKS pod, an ECS task or an EC2 workload in its own microVM with a guest kernel carved from the running instance by the Nitro hypervisor. Per AWS what's-new, it works on most Nitro-based EC2 instances in all commercial regions at no additional charge. AWS names coding agents first and positions the sandbox for safely executing LLM-generated code. Other scenarios: multi-tenant SaaS, regulated workloads, CI builds, remote IDEs.
How it works
On EKS, enabling the sandbox is one field in the pod spec: runtimeClassName set to isolated-sandbox or confidential-sandbox. Per EKS docs, the parent runs containerd-shim-nitro-sandbox-v2, which starts a Firecracker microVM with hardware-assisted virtualization, and an agent inside that VM runs containers through its own containerd over virtio and vsock. Sandboxes share neither kernel nor memory mappings with the parent or each other, containers start unmodified, and per what's-new Pod Identity, VPC CNI, storage and CloudWatch work out of the box.
Two modes, two trust models
Isolated mode trusts the parent, so kubectl exec, logs and process listing keep working. Confidential mode treats the parent as untrusted: a Cedar policy from the pod spec pins image digests, args, env, mounts and capabilities, and forbids exec and stdin. The Nitro Secure Module measures image and policy into PCRs, and per docs a KMS key policy can release secrets only to a sandbox with matching measurements. Per docs, without a KMS policy or external verifier the measurement is recorded but never verified, leaving confidential-mode guarantees incomplete.
Caveats worth planning for
Per InfoQ, AWS Container Hero Sandip Das notes that isolated-mode protection runs in one direction only: a compromised parent can see unencrypted traffic and DNS, rewrite /etc/hosts and mint Pod Identity credentials for any pod on the node. Prerequisites per Das: Kubernetes 1.36 or newer, Nitro Enclaves enabled in the launch template of managed node groups, otherwise sandboxed pods sit in Pending, plus Pod Identity instead of IRSA. Near-native performance is a vendor claim, and Das advises measuring startup latency, I/O throughput and tail latency yourself.
Limits and the wider picture
Firecracker cannot pass host devices into the guest, so GPU pods run without the sandbox class at launch. Total sandbox vCPUs and sandbox count stay at most 5x the node count: a 4 vCPU node fits twenty 1 vCPU sandboxes or five 4 vCPU ones, halved on CPU-overcommit types. With SMT, core scheduling forbids two sandboxes from sharing sibling threads of one core, leaving roughly half the vCPUs, though Graviton is unaffected. The kubernetes-sigs agent-sandbox serves a pre-warmed VM from warm pools in about 2 seconds, and GKE has offered a gVisor sandbox class since 2019 with starts under 200 ms.
Sources:
aws.amazon.com/blogs/containers/aws-nitro-sandboxes-hardware-enforced-workload-isolation-on-amazon-eks
aws.amazon.com/about-aws/whats-new/2026/11/nitro-sandboxes-eks-ecs-ec2
docs.aws.amazon.com/eks/latest/userguide/sandboxing.html
www.infoq.com/news/2026/11/aws-nitro-sandboxes-eks
Continue the conversation: https://t.me/AIAutomationArchitecture ยท Bluesky
Top comments (0)