Modern Workplace consulting · Microsoft cloud architectureAdelaide, Australia
October 2026 · Security

Microsoft 365 Copilot: review trust boundaries and prompt-injection risk

An October security review of untrusted content, human approval, agent permissions and practical AI pilot evidence.

Access permissions are an essential part of a Copilot rollout, but the security review also needs to consider how retrieved content influences the task. A document can be readable by the user and still contain instructions the assistant should not follow. October’s Microsoft security discussion provides a useful basis for reviewing that distinction.

The October security development

Microsoft’s 13 October article describes a trust-aware enforcement layer for Microsoft 365 Copilot. It uses origin signals to distinguish untrusted material, processes that material through a constrained path and passes a transformed representation onward. Initial investment focuses on higher-risk content such as external communications. Microsoft also points administrators to the applicable Outlook sender setting and associated security guidance, noting that availability and behaviour vary by release and tenant. [1]

This is a defence-in-depth development, not evidence that every prompt-injection attempt is prevented. Microsoft explains that indirect prompt injection can occur through content the user is already authorised to access. [1]

A practical review scenario

Consider a user asking for a summary of supplier correspondence. The business task is to extract relevant facts. A sentence inside a supplier email asking the assistant to change its task should be treated as source content, not as the user’s instruction.

For an approved pilot, use a synthetic email and a harmless conflicting instruction. Ask the system to summarise the message and record whether the response stays within the requested task. Do not include live customer records, real secrets or instructions that trigger external actions. This is a validation example, not a complete security test or a claim about a tested product result.

Review three boundaries separately

BoundaryReview questionEvidence
Content accessShould this identity be able to retrieve this information?Ownership, membership and permission review
Task interpretationDoes the assistant treat external material as information rather than authority?Approved synthetic test cases and observed outputs
Action executionWho authorises an agent to make a change or send information?Granted permissions, approval flow and action record

Keep these questions distinct in the design review. A successful permissions test does not establish that the assistant will interpret every document safely. A good answer to a prompt does not establish that an agent has appropriately restricted write permissions.

Treat agents as business services

For each agent, record its owner, permitted data sources, allowed actions and escalation path. Give read access and change authority separate scrutiny. Where an action creates a business commitment, define who reviews the proposed result before execution.

Start with a bounded use case such as preparing an internal draft from an approved document set. Establish acceptable output quality and review effort before expanding to additional sources or external interactions. Review abandoned pilots so unused permissions and integrations do not remain indefinitely.

Prepare the support and security response

Tell participants how to report suspicious or unexpected responses. Capture only the information needed for investigation, using approved handling for sensitive content. Record the task, relevant source, time and observed behaviour; avoid copying confidential prompts into an unrestricted ticket or public forum.

Before changing tenant configuration, follow the current Microsoft guidance linked in the source article and confirm applicability. My recommendation is to make one controlled change at a time, record the expected result and retest the same synthetic scenarios afterwards.

Pair this review with the Copilot adoption guide and the Frontier preview governance review. Together they provide a practical starting point for a pilot whose scope, evidence and ownership remain clear.

Microsoft references

  1. Microsoft: Building Trust-Aware Controls for AI, 13 October 2026

Sources checked 2 November 2026. Coverage focuses on October announcements through 31 October. Recommendations are proposed implementation practices; validate availability, licensing and requirements in your tenant before rollout.