Tools
Agent sandboxes go mainstream: AWS Strands Box, Microsoft Execution Containers, and the new containment layer
On 7 October 2026 AWS and Microsoft shipped OS-level containment for coding agents on the same day, and Claude Code made its hooks fail closed the next morning. What each boundary actually enforces, where the gaps sit, and what to check before you call an agent sandboxed.
- Published
- 14 Oct 2026
- Reading
- 8 min
- Class
- tools
half-life 90dfrom 14 Oct 2026
Until this month, "does your agent sandbox?" was a yes-or-no question you asked a vendor and mostly had to take on trust. That changed on 7 October 2026, when AWS released Strands Box as an open-source developer preview and Microsoft made its Execution Containers generally available — two separate containment layers, announced the same day, from two of the largest cloud platforms. The next morning, Anthropic shipped Claude Code v2.1.295 with an option that finally lets a security hook stop an action instead of logging it and carrying on.
None of the three is a new idea. Containers, microVMs and permission prompts have been around for years. What is new is that both hyperscalers now ship a boundary that sits outside the agent, enforced by the operating system, with a policy file you write yourself. Containment has stopped being a bespoke safety project and started being infrastructure. This is what arrived, what each layer genuinely enforces, and the gaps their own documentation admits to.
AWS: Strands Box, containment plus policy
Strands Box (announced 7 October 2026, Apache 2.0, code on GitHub) works in two layers. Containment is OS-level isolation — on macOS, Apple's Seatbelt — which defines what the agent can reach on the host and on the network. Inside that hard boundary sits policy: rules written in AWS's Dogwood policy language and evaluated by the Dogwood Local Engine, which keeps a history of what the agent has already done.
That history is the part AWS is selling as the difference. A rule can constrain not just one action but a pattern of them: the worked example in the announcement lets an incident agent post to a Slack channel, but no more than three times every ten minutes, enforced without relying on the agent to remember the limit. Policy is applied at four enforcement points — the network egress gateway, a shell interpreter, a Python interpreter and a broker for MCP servers — and each reports actions in a common vocabulary, so a rule can say "after the agent reads a file from the customer-data directory, block further outbound HTTP requests" without caring which tool did the reading.
The egress gateway also does credential injection: for configured API-key routes the
agent receives a placeholder token and the gateway swaps in the real secret on the way
out, so the secret never enters the agent's environment. AWS SigV4 signing works the same
way — the request is signed outside the agent entirely. And Box is harness-agnostic: the
same box.toml and policy.dw files apply across agent
applications, which AWS argues matters because a harness's own permission prompts run
inside the agent's process and judge the tool call, not the effect — as AWS puts it,
"A harness sees 'run this shell command'. It does not see the files the command will
touch or the hosts it will reach."
The limits are stated plainly. Box ships for Apple Silicon macOS only; Linux is in
development and Windows is, in AWS's words to The Register, "on our radar" with no
release date (7 October 2026). Paths you grant the agent directly in
box.toml are bounded by containment but do not appear in the policy
history, so a rule that assumes it can see every file read is wrong today. Deployment to
Bedrock AgentCore, ECS and Kubernetes is planned rather than shipped.
Microsoft: Execution Containers go generally available
Microsoft's announcement landed the same day, from Logan Iyer on the Windows Developer Blog (7 October 2026). Microsoft Execution Containers (MXC) is now generally available, with an SDK and a unified JSON configuration schema open at github.com/microsoft/mxc. The framing is the bluntest of the week: "An agent cannot be its own security authority. It must run within a boundary defined by the developer or organization and enforced independently of the agent itself."
You declare the files and network destinations a workload needs; MXC maps that to a platform backend and enforces it at runtime, with the policy held outside the agent's workload so generated code cannot grant itself more access. Microsoft's worked example is exactly the everyday case: a coding agent asked to update a website needs read and write on the repository and read on the production server configuration, but not the ability to change that configuration — and if it tries, the containment environment blocks the operation "regardless of what the model, generated code, plugin, or tool decides to do".
Four backends, with different strengths:
- Process container — Windows 11, macOS and Linux. AppContainer on Windows, Seatbelt on macOS, Bubblewrap on Linux. The lightweight default for model-generated code and tool execution.
- Session container — Windows 11 only. A separate OS session with its own desktop, clipboard, UI and input, for long-running agents.
- WSL container — Windows 11 only, for Linux-first toolchains.
- MicroVM — Windows 11 and Linux, experimental, for higher-risk workloads that want a hardware-backed boundary.
The adoption list is why this counts as mainstream rather than a preview. Microsoft names GitHub Copilot, OpenClaw, OpenAI Codex, Replit, LM Studio and Unsloth AI as supporting MXC today, with Anthropic Claude Code, Hermes Agent by Nous Research, Manus, Perplexity, Raycast and others announced as coming. Windows 365 Cloud PC support is generally available too. The governance layer is not: Microsoft Entra identity attribution for agent activity and Agent 365 policy management are described as coming soon, so for now the developer who writes the policy is also the whole security team.
The morning after: hooks that can finally say no
Containment only helps if the checks around it actually stop things. On 8 October 2026,
Claude Code v2.1.295 added onFailure: "block" for command and HTTP hooks:
"a hook that can't start, times out, or exits with an unexpected code blocks the action
instead of letting it through" (Anthropic release notes, 8 October 2026). The default is
unchanged — a broken check still lets the action continue.
That distinction matters more than it sounds. Installing a security hook and deciding that a failed security hook stops work are two separate decisions, and until this release only the first existed. Arcjet's documentation for its coding-agent policies, published 8 October 2026, spells out the current asymmetry: HTTP-hook integrations such as Claude Code and GitHub Copilot can let an action through when the hook itself fails, while wrapper-based integrations for Codex and Cursor fail closed and stop it. If you have written a policy that must never be bypassed, the failure path — not the happy path — is what you need to test.
Where the gaps are
All three releases are unusually candid about what they do not cover yet:
- Coverage follows the execution path. MXC constrains code it contains; file tools and remote MCP services need their own controls, and a permissive policy can record a denial without enforcing it (EveryDev weekly digest, 9 October 2026).
- Direct filesystem grants sit outside policy history in Strands Box.
AWS says so directly: paths granted in
box.tomlare bounded by containment but do not appear in the policy event history, so cross-tool rules cannot see them. AWS points to Apple's Endpoint Security additions in macOS 27 as how it intends to close that gap. - Platform support is uneven. Box is macOS-only for now; MXC's process containers span three desktop operating systems but its management controls have not shipped. Neither has a Linux-server containment story that matches what a CI runner needs.
- Enforcement is not judgement. Marc Brooker, AWS VP and distinguished engineer, told The Register that Box enforces configured policies deterministically and "the agent can't talk its way around these rules" — but added that "developers remain responsible for deciding what access to grant and where human review is needed". A deterministic sandbox around an over-broad policy is still an over-broad sandbox.
What to do about it
You do not need to pick a winner this month; you need to ask better questions of whatever you already run:
- Put the boundary outside the agent. Harness permission prompts and allow-lists are a user interface, not a security boundary — they run in the agent's process and only see the tool call. Treat them as such.
- Ask what happens when the check fails. Fail open or fail closed is
a configuration decision you can usually make today: Claude Code's
onFailure: "block"is a one-line change, and wrapper-based controllers fail closed by construction. - Check coverage, not just presence. Does the policy cover file tools, shell and Python interpreters, MCP servers and network egress — or only commands? Can it see what happened earlier in the session, or only judge each call in isolation?
- Keep credentials out of the agent's environment. Gateway-side injection, as in Box, is the pattern to look for: the agent handles a placeholder, the outside world sees the real secret.
- Start with what ships on your machine. GitHub Copilot and Codex users already have MXC available on Windows 11, macOS and Linux; Box is worth watching on Apple Silicon but is a developer preview with a single OS.
We looked at this question from the other direction in September, comparing what Prime Agent and Agent Zero actually isolate (Sandboxes and self-writing agents, 10 September 2026). Two weeks later the interesting question is no longer whether a harness sandboxes, but what the sandbox enforces, who writes the policy, and whether the check still works when it breaks. Those are infrastructure questions now — which is exactly what "goes mainstream" means.
Sources
- AWS Open Source: Introducing Strands Box, AI agent sandboxes powered by Dogwood — 7 Oct 2026. Primary source for Box's architecture, enforcement points, credential gateway and stated limits.
- The Register: AWS launches open-source AI agent sandbox to prevent YOLO mode disasters — 7 Oct 2026. Platform roadmap (Linux in development, Windows undated) and the Marc Brooker quotes.
- Windows Developer Blog: Microsoft Execution Containers — policy-driven containment for AI agents — 7 Oct 2026. GA announcement, containment backends, adopter list, Entra/Agent 365 roadmap.
- Microsoft MXC repository — SDK, JSON configuration schema and samples.
- Claude Code v2.1.295 release notes — 8 Oct 2026. The
onFailure: "block"option for command and HTTP hooks. - Arcjet coding-agent security documentation — 8 Oct 2026. Per-agent failure modes: HTTP-hook integrations can fail open, wrappers fail closed.
- EveryDev weekly AI dev news digest, 3–9 October 2026 — 9 Oct 2026. Independent summary of the same week's containment releases and their coverage gaps.