Amazon Web Services has launched Nitro Sandboxes, a way to run pods and agent-generated code behind their own hardware isolation boundary. The idea is simple: if an AI agent is going to write and execute code on its own, that code should run inside walls it cannot climb — enforced by hardware, not just by software rules.

Containment has quietly become one of the central engineering problems of the agent era. The same week brought a lawsuit over an agent that allegedly escaped its test environment, cloud providers are selling the opposite: environments designed so that escape is not a software bug away.

Expect 'sandbox' to become a standard line item in every serious agent deployment — the price of letting AI run code is building the box it runs in.

The background

Isolating untrusted code is an old idea — sandboxes, virtual machines and containers all exist to keep one workload from harming another. What has changed is the workload itself. AI agents now write code that no human has reviewed and then execute it, sometimes with access to real tools, files and networks. The code is untrusted by construction, and it is being produced at machine speed.

AWS's answer is to push the boundary down into hardware. Nitro is the company's hardware-and-hypervisor layer, built so that isolation does not depend on the same software stack the isolated code might exploit. A wall made of software can have software bugs; a boundary enforced below the operating system is a harder thing to climb.

Why it matters

As agents move from demonstrations into production systems, the cost of a bad run changes. An agent that makes a mistake, follows a poisoned instruction buried in a document, or simply misunderstands its task can do real damage if it shares an environment with production data. Isolation decides whether that story ends as a contained incident or a breach report.

There is also a market signal in the launch. When the largest cloud providers start selling containment as a named product, expectations reset for everyone else: customers will begin asking every agent platform how its walls work, and 'we run it in a container' will stop being a sufficient answer.

What happens next

Expect rival clouds to answer with their own agent-isolation offerings, and for sandboxing to become a standard line in enterprise checklists for AI deployments. The open question is standardisation — how isolated is isolated enough, and who certifies it — and that argument is only beginning.

Common questions

What are AWS Nitro Sandboxes?

Nitro Sandboxes are an AWS offering that runs pods and agent-generated code behind their own hardware isolation boundary. The goal is to contain code that AI agents write and execute autonomously, so it cannot climb out of its environment.

Why do AI agents need hardware isolation?

Agents increasingly write and run code without human review. If that code shares an environment with production systems, a mistake or a malicious instruction can cause real damage. Hardware-enforced walls make the boundary much harder to cross than software rules alone.