Microsoft Execution Containers Hit GA: What Agent Isolation Misses
At its October 7, 2026 Windows event, Microsoft moved Microsoft Execution Containers, which isolate AI agents, from preview to general availability. The idea is sound: when an agent runs on your machine, put it in a sandbox so it cannot freely touch the rest of your system. Isolation is a real security control and a welcome default. It is also easy to overread, so it is worth separating what a container protects from what it leaves exposed.
What A Container Actually Does
An execution container limits an agent’s reach. Inside the sandbox, the agent can do its work; outside it, the system constrains what files, processes, and resources the agent can access. If the agent misbehaves or is compromised, the blast radius is bounded by the container rather than your whole machine.
That is genuinely valuable. An agent running arbitrary actions on your system is a large attack surface, and bounding it is the correct engineering response. For agents that execute code or move files, a sandbox is close to mandatory, and making it the default is the right call.
What It Does Not Do: The Data You Hand It
Here is the first gap. A container controls what the agent can reach on your system. It does nothing about the data you deliberately give the agent to work with.
If you hand an agent a confidential document to summarize, the sandbox does not stop that document from being sent to a cloud model, logged, or retained, because you authorized the agent to process it. Isolation is about the agent’s access to your system, not about where your input travels once it is inside the agent’s legitimate workflow. The broader version of this distinction is in what a secure VM AI agent does not protect.
What It Does Not Do: Acting Within Allowed Scope
The second gap is subtler and matters more. A container bounds what an agent can do. It does not judge whether the agent should do a particular thing within that boundary.
If an agent is allowed to send email, and a web page it reads contains a hidden instruction telling it to send your data somewhere, a container that permits email does not stop that send. The action is inside the allowed scope; the problem is that the agent was manipulated into choosing it. That is prompt injection, and sandboxing is not a defense against it, as explained in indirect prompt injection for non-engineers. What your data is exposed to the moment an agent acts is covered in what your data is exposed to when an agent acts.
A Sandbox Is Not Local AI
It is tempting to treat “runs in a container on my machine” as equivalent to “stays on my machine.” It is not. An agent in a local sandbox can still reach the network and send your data to a cloud service; the container limits system access, not egress. The difference is laid out in a sandboxed agent is not the same as local AI.
Local AI makes a narrower, stronger promise about one operation: the rewrite runs on your machine and makes zero network calls, so the text does not leave to be processed. A container makes no such promise, and conflating the two is how people end up surprised.
Before:
The agent runs in an isolated container, so our confidential drafts are safe to hand it.
After:
The container limits what the agent can touch on our system. It does not control where a document we hand the agent travels, or stop the agent from being tricked into a permitted action. We will redact sensitive drafts before the agent sees them.
The second version treats isolation as one layer, not the whole defense. A Wrivio Context for preparing text before an agent handles it could say:
Rewrite this so it is safe to hand to an automated agent. Remove confidential figures, internal names, and anything the agent does not need to complete the task. Keep every instruction, date, and commitment that must remain exactly as written. Do not add new instructions or steps.
Press Ctrl+Shift+Space, paste the draft, and check the diff before the agent touches it. Microsoft’s own documentation on Microsoft Learn is the place to confirm what the container does and does not scope; treat these details as current to October 2026. For deciding when an agent should act at all, see when to let an agent act and when to only draft.
Common Questions
What are Microsoft Execution Containers?
A feature that moved to general availability in October 2026 that isolates AI agents in a sandbox, limiting what files, processes, and resources an agent can reach on your system so a compromised or misbehaving agent has a bounded blast radius.
Does a container keep my data private?
It controls the agent’s access to your system, not where the data you hand the agent travels; a document you give an agent to process can still be sent to a cloud model or logged within the agent’s legitimate workflow.
Does isolation stop prompt injection?
No, a container bounds what an agent can do but does not judge whether a specific permitted action is a good idea, so an agent manipulated into a send it is allowed to make will still make it.
Is a sandboxed agent the same as local AI?
No, a local sandbox limits system access but the agent can still reach the network and send your data to the cloud, whereas local AI means the operation runs on your machine with zero network calls.
Download Wrivio for Windows to redact a sensitive draft on-device before any sandboxed agent gets near it.
Read Next
Air-Gapped AI Writing: What It Really Means
Air-gapped sounds absolute, but most tools that claim it are not. What air-gapped actually requires, how local AI gets close, and how to verify the claim.
How to Redact a Document Before an AI Rewrite
If you must use a cloud AI on sensitive text, redact first. What to strip, how to keep the draft useful, and why local AI skips the problem entirely.
Agentic Browsers And The Same-Origin Problem: Why Your Draft Is Exposed
A 2026 University of Washington study found four of seven agentic browsers could be tricked into reading across sites. What that means before you paste confidential text.
What To Do When A Local Model Stalls Mid-Rewrite
A local rewrite that hangs or crawls has a handful of ordinary causes. How to tell which one you have and fix it, without blaming the model or your machine.
This article is filed underPrivacy & Compliance, which has 114 articles.