Insights

Windows AI security: what to check before enabling local agents

29 October 2026

Review Windows AI security, device readiness, agent permissions and rollout testing before enabling local AI tools across your business devices.

Microsoft has outlined changes intended to make Windows a more controlled platform for AI agents, according to Computer Weekly’s report. For Australian businesses, the Windows AI security question is practical: which devices can support the tools, what can each agent access, and how will IT prove the controls work before deployment?

The report describes Microsoft Execution Containers (MXC) as generally available, with controls intended to restrict agents’ access to files and business resources. That is a reason to investigate, not permission to enable agents across every workstation.

What the reported Windows changes mean

Computer Weekly attributes the announcement to Microsoft and describes a direction built around containment, agent identity and central management. Those are useful foundations: an agent should have a defined operating boundary, an identifiable account or execution context, and controls administrators can manage.

However, the supplied report is not a deployment specification. It does not establish the supported Windows builds, editions, hardware requirements, licensing or exact policy settings for your environment. Its description of update timing should not become your deployment schedule.

Before planning a rollout, obtain Microsoft’s current feature documentation and check the relevant build against Windows release health. Confirm support with the agent vendor as well. A Windows capability being available does not establish that every application uses it correctly or supports the same controls.

Keep three statuses separate in your planning: available now, available only in preview, and planned integration. Record the documentation date and tested version rather than relying on an announcement headline.

Windows AI security starts with device readiness

Start with an inventory, not a purchase order. Windows AI security depends on a supported, managed endpoint as well as the agent itself.

For each proposed pilot device, record:

  • Operating system: Windows edition, build, update status and support lifecycle.
  • Compute capacity: processor architecture, memory, available storage and any GPU or NPU requirement documented for the selected workload.
  • Security configuration: encryption, Secure Boot, endpoint protection and any documented virtualisation prerequisites.
  • Management coverage: whether IT can apply policies, collect relevant events and remove the application remotely.
  • Application dependencies: required runtimes, model downloads, connectors, drivers and update mechanisms.

Do not assume that every local agent requires an NPU, or that an AI-branded laptop will support every model. Requirements depend on the application, model and execution method.

Test on representative equipment while ordinary business applications are running. A successful demonstration on a high-specification machine says little about a finance laptop handling spreadsheets, calls and invoice software simultaneously. Measure responsiveness, storage growth, battery impact and stability before approving broader use.

Define permissions around the task

Containment can help limit access, but it does not decide which access is appropriate. Treat Windows AI security as a permissions design exercise, not just an installation task.

Write down the intended workflow in plain language. For example: read documents from one approved folder, extract specified fields and save a draft to a review location. That description is easier to turn into an access policy than a broad instruction to “help with finance”.

Check these boundaries explicitly:

  • Files: separate approved inputs and draft outputs from payroll, credentials and unrelated customer records.
  • Network access: identify permitted destinations, including model services, application APIs and update endpoints.
  • Identity: use a dedicated, restricted identity where supported; do not hand an agent an administrator’s session or reusable credentials.
  • Actions: distinguish reading and drafting from changing records, sending messages, deleting files or initiating payments.
  • Approval: require a person to review consequential actions before execution, with a clear record of what they approved.

An agent may reach business data through a browser session or connector even when local folders are restricted. Review those routes too. Our AI agent identity security checklist explains why accounts, permissions and approvals need to be designed together.

Local execution does not settle data handling

Running a model on a workstation may reduce some external processing, but “local” is not a complete data-flow description. The application may still contact cloud services for authentication, telemetry, updates, search or fallback inference.

Ask the vendor to document what leaves the device, where it goes, how long it is retained and whether administrators can control those transfers. Then compare that documentation with observed behaviour during testing.

Also check what remains behind. Prompts, extracted text, model caches, temporary files and diagnostic logs can contain sensitive material. Establish who can access them, how they are protected and how they are removed when a device is reassigned or retired.

A Windows containment feature should not be treated as proof of Australian data residency or regulatory compliance. Those conclusions require evidence about the full application and its connected services.

Test the controls before expanding the rollout

A useful Windows AI security pilot tests failure as deliberately as success. Use synthetic records or suitably de-identified data, a limited device group and a named business owner.

Include the following checks in the acceptance plan:

| Test | Evidence to capture | | --- | --- | | Read an approved document | The task completes using only authorised inputs | | Request a restricted file | Access is denied and the attempt is visible to the administrator | | Place malicious instructions inside a test document | The agent does not disclose data or perform an unauthorised action | | Request a consequential change | The workflow pauses for human approval before execution | | Remove permissions during a session | Access stops within the documented enforcement behaviour | | Disconnect the network | Any failure or fallback is understood and does not bypass policy | | Disable the agent | IT can stop execution, revoke access and identify unfinished work |

Also verify that logs can connect the request, agent identity, resource access, output and approval without unnecessarily recording sensitive content. A generic “task completed” event is not enough to investigate an incorrect change.

Keep a tested recovery path for affected files and business records. Removing software does not undo actions it has already taken. For broader acceptance criteria, use our guide to AI governance controls businesses can test.

Separate local Windows agents from managed business agents

Local Windows AI and Tech Engine’s managed AI agents have different execution environments. Local tools depend on endpoint capacity, device state and the application’s Windows integration. Tech Engine agents run with human approval in a private Australian-hosted environment and integrate with existing business systems.

For example, an Order and Data Intake Agent can prepare validated records from incoming orders. An AP/AR Automation Agent can match supplier invoices to records and prepare draft customer invoices. These workflows keep human approval rather than giving a workstation agent unrestricted authority over business transactions.

That distinction does not make endpoint security irrelevant. Staff still need secure devices and controlled access to connected systems. It does mean that running the agent does not depend on installing a local model on every employee’s computer.

Choose the execution environment around the workflow, data boundaries and operational responsibilities—not the availability of a new Windows feature.

Make deployment an owned operational decision

Before expanding beyond the pilot, assign responsibility for policy changes, application updates, model updates, access reviews and incident response. Changes to the agent or its connectors should trigger proportionate retesting, even if Windows itself has not changed.

A sound Windows AI security decision has documented support requirements, tested access boundaries, working approvals and a way to stop and recover the workflow.

Tech Engine can help assess device readiness and design managed AI agents that integrate with your existing systems, with human approval in a private Australian-hosted environment. Discuss your workflow at aiaas.au, call 1300 088 324 or email sales@techengine.au.

Want this applied to your business?

Request an AI Blueprint and we will map the processes worth automating first.

Request an AI Blueprint