AWS Nitro Sandboxes Are Betting Hardware Beats Software

The headline framing around AWS Nitro Sandboxes is "containers get their own VM." The real story is narrower and more interesting: AWS just admitted that letting an AI agent run code it wrote itself is now a mainstream enterprise problem, serious enough to solve with silicon instead of software. That shift, from trusting a container to trusting a chip, says more about where agentic AI is headed than any benchmark release this quarter.

Key takeaways

  • AWS Nitro Sandboxes launched November 9, 2026, giving each Amazon EKS pod, ECS task, or EC2-hosted workload its own hardware-isolated VM, with AWS explicitly naming "AI agents that run code from their users" as a target use case.
  • There is no additional charge for the feature itself; you only pay the normal EC2, EKS, or ECS bill, according to AWS's own EC2 User Guide.
  • GPU passthrough into a sandbox is not supported yet, so on accelerated computing instances a Nitro Sandbox can only run CPU workloads, a real limit for agents that need in-sandbox inference.
  • AWS is following, not inventing, a trend: open-source Kata Containers, AWS's own Firecracker (already behind Lambda), and Google's gVisor all solve the same untrusted-code problem with lightweight VM or kernel-level isolation.

AI Agents Running Code Broke the Old Sandbox Assumptions

For most of the container era, the thing running inside a sandbox was code someone on your team wrote and reviewed. The isolation layer, whether it was a container's namespaces or a lightweight VM, existed mainly to contain bugs and limit blast radius, not to defend against an adversary sitting inside the box. That assumption is what agentic AI breaks. When a coding agent generates a shell command or a data-processing script and executes it immediately, the code running inside your infrastructure was never reviewed by a human, and sometimes it is shaped by a user's prompt rather than your own engineers.

That is already visible in how AI tooling is evolving. Browser-automation tools like Chrome DevTools MCP letting AI agents write files directly and the broader practice of AI labs running agent tests against the live internet both point to the same pressure: agents need real execution environments to be useful, and real execution environments need a much harder isolation boundary than a typical container gives you.

Advertisement

What AWS Actually Shipped on November 9

Nitro Sandboxes is an Amazon EC2 feature, not a separate product. A Nitro Sandbox is a lightweight virtual machine that runs alongside a normal EC2 instance, called the parent instance, drawing its vCPUs and memory from that same host. A virtual machine monitor, currently only Firecracker, creates the sandbox through the Nitro Enclaves driver, and the Nitro hypervisor, the same layer that has isolated EC2 instances from each other for years, enforces the boundary around it, according to AWS's EC2 User Guide.

AWS ships two trust modes, and the choice is permanent once you launch a sandbox. Confidential mode, the default, hides the sandbox's memory from the parent instance and its administrators entirely; a Nitro Security Module inside the sandbox produces a signed attestation document, and AWS Key Management Service will only release a secret if that document matches your key policy. Direct mode drops the attestation and lets the parent instance read and write the sandbox's memory directly, which AWS describes as equivalent to a nested VM's trust model. The feature plugs natively into orchestration: Amazon EKS runs each pod in its own Nitro Sandbox, and Amazon ECS does the same for each task, so most users will never touch Firecracker directly.

A developer's screen showing lines of code

The Capacity Numbers Reveal the Real Design Target

AWS's own limits are the clearest evidence of who this is built for. A single parent instance can host up to 896 Nitro Sandboxes, and with CPU overcommit enabled, the combined vCPUs across all of a host's sandboxes can reach five times the parent instance's own vCPU count. Nobody designs for 896 isolated VMs per host to run a handful of long-lived internal services. That ceiling is built for high-churn, short-lived sessions: one sandbox per agent conversation, one per build job, one per tenant request, spun up and torn down constantly.

Advertisement

That density target also explains the pricing decision. Charging per sandbox would have made agentic workloads, which can spin up dozens of short sandboxes per task, unpredictably expensive. By folding the cost into the existing EC2, EKS, or ECS bill instead, AWS removes a line-item objection at exactly the moment enterprises are trying to figure out what running autonomous agents at scale will actually cost them, a question that is already colliding with the per-token pricing fights among AI model providers.

✦ Free Newsletter ✦

Never miss a story

Tools, tutorials and AI deep-dives - straight to your inbox, every week.

No spam, unsubscribe any time.
Macro shot of a computer circuit board

This Is an Industry Pattern, Not a Single Vendor's Idea

AWS is not the first to reach for VM-level isolation instead of trusting container namespaces alone. Firecracker, the same microVM technology underneath Nitro Sandboxes, is the technology AWS already built years ago to isolate Lambda functions and Fargate tasks from each other on shared hardware. Google took a related but different path with gVisor, a sandboxed kernel that intercepts a workload's system calls instead of handing it a full virtual machine, and uses it to isolate Cloud Run and GKE Sandbox workloads. The open-source Kata Containers project, built on Apache 2.0 licensing, runs containers inside lightweight VMs with their own guest kernels too, but deliberately supports multiple hypervisors, including QEMU, Cloud Hypervisor, and Firecracker itself, according to a CloudNativeNow report on the Nitro Sandboxes launch.

The pattern across all three approaches is the same conclusion reached from different directions: once the thing inside the box might be actively hostile rather than merely buggy, process-level isolation stops being enough, and you need either a separate kernel or a hardware-enforced boundary. AWS's director of security, Mark Ryland, framed the company's version of that philosophy bluntly in comments reported by CloudNativeNow, describing Nitro's approach as stripping "the hypervisor down to the absolute minimum" and moving everything else onto separate, dedicated hardware.

A padlock resting on a computer keyboard

The Honest Limits: No GPUs, No Live Migration, No Portability

AWS's own documentation undercuts some of the launch excitement. GPU passthrough into a Nitro Sandbox is not available, and on accelerated computing instances a sandbox can only run CPU workloads. That matters because a growing share of agentic workloads want to run inference or lightweight model calls inside the same isolated boundary as the generated code, and today that is simply off the table on Nitro Sandboxes. AWS's documentation also confirms sandboxes cannot be snapshotted, paused and resumed, or live-migrated, which rules out several operational patterns that platform teams rely on for ordinary VMs.

Advertisement

There is also a fair counterargument to raise here. Someone building on AWS could reasonably argue that tying isolation to the Nitro hypervisor, rather than to a portable open-source runtime, is a feature rather than a flaw: the attack surface is smaller when the boundary is enforced by hardware AWS controls end to end, instead of a software layer running across arbitrary hypervisors. That argument has real merit, and it is the same logic behind AWS's confidential-computing pitch more broadly. But the tradeoff is real too: Kata Containers' multi-hypervisor design means a team can move the same isolation guarantee between AWS, another cloud, or their own data center, while Nitro Sandboxes only exists on supported Nitro-based EC2 instances. You get AWS's strongest guarantee only by accepting AWS's strongest lock-in.

Abstract visualization of cloud computing infrastructure

Who Should Care About This, and Who Can Ignore It

This matters most to platform and security teams already running on EKS, ECS, or EC2 who are building or operating agents that execute model-generated or user-submitted code, and to anyone running genuinely multi-tenant workloads on shared instances. For that audience, Nitro Sandboxes is a meaningful, no-extra-cost upgrade over bolting software sandboxes onto containers by hand.

  • Use it if: you are already on AWS, your agents or multi-tenant jobs run on EKS, ECS, or EC2, and hardware-enforced isolation with optional attestation solves a real threat model you have today.
  • Skip it if: you need portability across clouds or on-prem infrastructure, your agents need GPU access inside the isolated boundary right now, or your workloads depend on snapshotting, pausing, or live-migrating running sandboxes.
  • Check first: your instance family is actually supported. Flex instances such as C7i-flex and M7i-flex are explicitly excluded, and accelerated instances only get CPU-side isolation.

What to Watch Next

The feature to watch for is GPU passthrough. Once AWS can isolate GPU-backed inference the same way it now isolates CPU execution, Nitro Sandboxes becomes a far more complete answer for agents that both generate and run model-backed code in one place. Also watch whether Microsoft and Google answer with their own hardware-anchored pitch for agent isolation specifically, rather than leaving the positioning to general confidential-computing marketing; gVisor and Azure's own isolation layers already have the technical pieces, they just have not been marketed at "AI agents running your users' code" the way AWS now has. For now, the safest reading is that AWS correctly identified a real and growing problem, built a reasonable answer to it, and priced that answer to make switching away from AWS a little harder than it was last month.

Sources

Joe Manning
Written by
Joe Manning, Senior Editor
Share this article:
Advertisement