← Back to Resources
Open sourceOctober 27, 2026

Strands Box: AWS Puts a Cage Around AI Agents

Strands Box is AWS's answer to a question every team running agents eventually has to face: what happens when an agent holding real credentials meets a prompt it was tricked into misreading? Announced as a developer preview on the AWS Open Source Blog and published as open source under the strands-agents organization on GitHub, Strands Box wraps agent workloads in an OS-isolated sandbox, gates every action through a default-deny policy engine named Dogwood, inspects MCP calls and outbound traffic, and injects credentials only when a policy allows it.

What AWS shipped

The announcement came through the AWS Open Source Blog as a developer preview titled around "AI agent sandboxes powered by Dogwood." Two facts matter more than the marketing copy. First, it is open source and lives in the same strands-agents organization as AWS's Strands agent framework, so it is positioned as infrastructure for the agent ecosystem rather than a proprietary add-on. Second, it is explicitly a preview: the policy format and interfaces are still moving, and AWS is asking for feedback rather than declaring a finished product.

The preview bundles four capabilities that usually live in separate tools: an OS-isolated execution environment, the Dogwood policy engine, inspection of MCP traffic and network egress, and just-in-time credential injection. Together they address the same failure mode from four angles—an agent doing something it should not, somewhere it should not, with secrets it should not have.

Why agent sandboxes exist at all

A traditional application has a fixed call graph: developers decide what it can touch, and reviewers check that decision once. An agent decides at runtime which tools to call, which files to read, and which URLs to fetch, based on input that may include untrusted content. That flexibility is the product—and it is also the attack surface. A prompt injected through a web page, an issue comment, or a PDF can redirect a helpful agent into exfiltrating files, spending money, or writing to production systems.

Most teams respond with a patchwork: a system prompt asking nicely, a code review of the tool list, and hope. Sandboxes replace hope with mechanism. The question shifts from "will the model behave?" to "if the model misbehaves, what can it actually reach?" Strands Box is AWS's attempt to make the second question easy to answer with configuration instead of incident response.

Dogwood: default-deny by design

The most consequential design decision in Strands Box is Dogwood's default posture. Security tooling for agents often starts from default-allow: everything works until you write a rule that blocks something. Dogwood inverts that. By default, everything is denied—tool invocations, MCP server connections, outbound network destinations. An explicit policy grants each capability, and anything the policy does not mention fails closed.

Default-deny has a cost: getting a complex agent running means enumerating what it legitimately needs, which is genuinely tedious. The payoff arrives the first time a model hallucinates a tool name or a prompt injection tells it to POST your environment to an unknown domain. Those attempts hit a policy wall instead of a live endpoint. It also changes the review conversation—policies are declarative artifacts you can diff, audit, and version, not vibes embedded in a system prompt.

Under the hood: isolation, MCP checks, egress, credentials

  • OS-level isolation. Agent processes run in a sandbox enforced by the operating system, so a runaway or compromised agent cannot freely read the host filesystem or spawn arbitrary processes.
  • MCP inspection. Model Context Protocol traffic is checked against policy per server and per tool, which matters because MCP has become the standard way agents reach external systems—and therefore the standard way attacks try to reach them.
  • Egress checks. Outbound network calls are filtered, so even an agent that has been manipulated cannot phone home to an exfiltration endpoint the policy never named.
  • Credential injection. Secrets are delivered just in time, when a policy permits an action, rather than sitting in the agent's environment where one successful prompt injection could read them all.

The combination is the point. Any one control can be engineered around; together they shrink the blast radius of a compromised run to roughly what the policy author intended.

How the common approaches compare

ApproachWhat it controlsWhat it misses
No sandbox, prompt rules onlyNothing enforceable—requests the model behaveEverything; a single injection defeats it
Generic containerFilesystem and process isolationTool-call policy, MCP traffic, network egress, standing secrets in env vars
Harness-level guardrailsTool-call allowlists, approval prompts in your own agent codeOS-level escape, direct network access, config drift
Strands Box (preview)OS isolation plus Dogwood default-deny policy, MCP and egress checks, injected credentialsMaturity—it is a developer preview, not a hardened product yet

Where Strands Box fits in an agent stack

Sandboxing is one layer of several. At the harness level, tools like Agent Tool Call Guard validate tool invocations before they fire, and WithHuman puts a human approval step in front of high-risk actions. For agents that purchase things, Composio Instant treats spend as a first-class, wallet-scoped permission. A sandbox like Strands Box complements these rather than replacing them: the harness decides what the agent may intend, the sandbox decides what it can physically reach.

It also arrives in the same fortnight as a broader push to treat agents as engineered systems rather than demos—see our pages on the Cursor Agent SDK and Anthropic's single-agent loops for the development-side counterpart.

Getting started without betting production on a preview

  • Clone the repo and read the example policies before writing your own—the format is the product.
  • Point one internal, low-stakes agent at it. A run that fails closed teaches you more than one that succeeds open.
  • Log every denied action during the trial. The denial log is a free audit of what your agents actually try to do.
  • Watch the preview's changelog for policy-format changes before committing your rules to version control pipelines.

Strands Box FAQ

What is Strands Box?

Strands Box is an open-source sandbox for AI agents from AWS, announced as a developer preview on the AWS Open Source Blog and published under the strands-agents organization on GitHub. It runs agent workloads in an OS-isolated environment, enforces behavior through a default-deny policy engine called Dogwood, inspects MCP calls and outbound network traffic, and injects credentials only at the moment a policy allows them.

What is Dogwood?

Dogwood is the default-deny policy engine behind Strands Box. Instead of blocking a list of known-bad behaviors, it denies everything by default and grants access—tool calls, MCP servers, outbound network—only when an explicit policy allows it. For agents that touch real systems, default-deny is the safer posture because behavior you never anticipated fails closed rather than open.

How is Strands Box different from running an agent in Docker?

A plain container isolates the filesystem and process space, but it usually says nothing about which tools an agent may call, which MCP servers it may reach, or whether credentials leak out through network calls. Strands Box layers a policy plane on top of isolation: OS-level sandboxing plus per-call inspection of MCP traffic and egress, with credentials injected at use time instead of baked into the environment. Containers remain a building block—Strands Box is the control layer wrapped around them.

How do I get started with Strands Box?

Start with the repository at github.com/strands-agents/box and the AWS Open Source Blog announcement, which introduces the developer preview and the Dogwood model. Read the policy format before wiring anything in, then run a low-risk workload—a scraper or a doc-summarizing agent—under a policy that denies everything except what you explicitly allow. Treat the preview as a chance to influence the policy model, not as a turnkey production control plane.

Is Strands Box production-ready?

No. AWS labels it a developer preview, which means the APIs and policy format can change, platform coverage may be incomplete, and you should not route production traffic through it yet. The sensible move is to evaluate it against your threat model now, note where the policy engine falls short for your workloads, and keep current guardrails in place until the preview matures.

AIXList is an independent AI tools directory. This article summarizes Strands Box based on the AWS Open Source Blog developer-preview announcement and the public GitHub repository as of October 27, 2026; we are not affiliated with AWS. Capabilities, policy formats, and platform support may change—verify against the official blog post and repository before building on it.