← Back to blog

AI Agent Access Control in Microsoft 365: A Practical Guide

November 14, 2026
AI Agent Access Control in Microsoft 365: A Practical Guide

AI agent access control means knowing which agents exist, who created them, what they connect to and what each one is allowed to do, then proving that the gaps you find actually get closed. In a Microsoft 365 tenant, agents appear faster than any quarterly review can follow, so the control that matters most is a daily, evidence-based audit that ties every agent to a named person. That is what Argus does for Microsoft 365: it audits all AI agents daily, records creators, connection points and permissions, ranks the risks and verifies the fixes.

This guide is for security teams, compliance officers, business managers and CISOs who have to answer one question: "Which agents can reach what, and who is accountable for each?"

Secureai
secureai.ie
See Every Microsoft 365 AI Agent
Argus audits AI agents daily, showing their creators, connections and permissions, and verifies fixes to help teams manage access risks.
Visit

TL;DR

  • An agent is a new kind of identity. It often holds more access than the person who built it, and it acts without anyone reading each request.
  • Start with an inventory that links every agent to its creator, current owner, connections and permissions. A static register goes stale quickly.
  • Apply least privilege: the minimum access, for the shortest time, with human approval for high-impact actions.
  • Rank risks (unrestricted sharing, sensitive connections), assign each to an owner and ticket it.
  • Closing a ticket is not proof of a fix. Re-check the agent and keep the evidence for auditors.

Table of Contents

Why agents need more than ordinary user permissions

A user permission model assumes a person who reads a screen, understands why a rule exists and can be trained. An agent does none of that. It works through connectors, APIs and data sources, and it makes many decisions in a single session.

Ireland's National Cyber Security Centre (NCSC) addresses this directly in its 2026 AI Cyber Security Risk Assessment. It notes that AI agents often hold elevated credentials and connect to databases, APIs and operational tooling. A compromised agent can then take unauthorized actions, move laterally and exfiltrate data. The same assessment describes "excessive agency": an agent with more functions, permissions or autonomy than it needs. The cure is not a smarter agent. It is a narrower one.

Three practical differences follow for a Microsoft 365 tenant:

  1. Agents are created by many people. Makers in any department can build one, so the population is not limited to what IT deployed.
  2. Agents borrow authority. An agent may run with its maker's credentials or the end user's. Which one applies decides how much it can reach.
  3. Agents are shared. An agent that is open to everyone in the organization, or beyond, extends its access to every person who can talk to it.

Microsoft documents the second point for Copilot Studio. Administrators can control maker-provided credentials so that each end user only reaches what their own account allows, instead of everything the maker could reach. That is a platform setting you need to know about. Knowing which of your agents still rely on maker credentials is the audit question.

Step 1: Build an inventory that survives contact with reality

You cannot control access for agents you do not know exist. Analysts at KuppingerCole argue that effective discovery needs broad coverage of sanctioned and unsanctioned agents, identity context and lifecycle monitoring, kept as a continuously updated picture and not a static list. They add that a quarterly register is obsolete, because agents are created, modified, duplicated and retired much faster than traditional applications.

For each agent, the inventory should record:

  • the creator and the current owner (these often differ after someone changes role or leaves)
  • the connections it uses: connectors, flows, knowledge sources, other services
  • the permissions and sharing scope: who can use it and who can edit it
  • the credentials it runs under
  • when it was created and last changed

Doing this by hand across a tenant means exporting data from several admin centers and reconciling it in a spreadsheet that is out of date the next morning. Argus runs the audit every day and records creators, connection points and permissions for all agents in the tenant. It links each agent to its creator and current owner, so the register is a live record and not a snapshot. It works without making changes to your tenant and without requiring additional Microsoft licenses, so the inventory step does not create a new project of its own.

Step 2: Decide who is accountable for each agent

Access control without ownership is a list of permissions nobody defends. The UK NCSC's June 2026 guidance, Thinking carefully before adopting agentic AI, says to define before connection who owns the system, who approves its access, who monitors its behavior, who reviews incidents and who can stop it. France's cybersecurity agency ANSSI asks for the same in its guide Securing an Agentic AI System: an up-to-date register of the people responsible for the system's policies, alerts and exceptions, revised after any change or incident and at least annually.

For a compliance officer, this is the heart of the matter. When an auditor asks who approved an agent's access to a sensitive SharePoint library, "the marketing team" is not an answer. A named person with a date is.

A workable ownership model has three rules:

  • Every agent has one accountable owner, recorded where security can see it.
  • Ownership transfers when the owner changes role or leaves, not when someone notices.
  • Orphaned agents (no active owner) are treated as findings, ranked and ticketed like any other risk.

Argus ties each agent to its creator and current owner, which is the record regulated organizations need to answer accountability questions. Because it re-checks daily, an agent whose owner has left shows up as a finding the next day, not at the next annual review.

Step 3: Apply least privilege to every agent

Least privilege is the principle that carries across every official source on this topic. The UK NCSC recommends minimum access for the shortest time, temporary credentials with elevated access revoked when a task completes, and monitoring for unusual activity across tools, workflows and connected systems. It also says never to give an agent unrestricted access to sensitive data or critical systems.

Server racks in an Irish data centre

ANSSI is more prescriptive. It asks organizations to inventory each agentic system's accesses, resources and capabilities, justify that each one is strictly necessary, and re-document the system's mission at least annually or after any significant change. It also asks for dedicated identities, secrets and connectors per environment (development, evaluation, production) and short-lived, least-privilege tokens scoped to operational needs.

Finland's cybersecurity authority Traficom adds a design-stage point in its guide Cybersecurity of AI agents: potential conflicts in data access rights should be identified when the agent is designed, especially where it has broad access to data.

A permissions checklist you can apply this week

QuestionWhat a good answer looks like
Whose credentials does the agent use?The end user's, so each person only reaches what their own account allows
Who can use the agent?A defined group, not the whole organization by default
What does it connect to?Only sources the task needs, with sensitive ones justified in writing
Who can edit it?The owner and a named backup
What can it do without a human?Read and summarize; anything irreversible needs approval
When was access last reviewed?Within the last year, and after every significant change

The last two rows matter most. ANSSI asks for human approval of critical agent actions, including access to sensitive data, changes to identity or permissions and production deployments, with the approval mechanism kept outside the agent's reach. An agent that can approve its own escalation is not under control.

Step 4: Rank the risks, so the team fixes the right ones first

A tenant with hundreds of agents will produce hundreds of findings. A list that long gets ignored. What the team needs is a ranking.

In Microsoft 365, two patterns deserve attention first:

  • Unrestricted sharing. An agent that anyone in the organization can use inherits the reach of its connections for all of them. The wider the audience, the more the agent's access matters.
  • Sensitive connections. An agent wired into HR, finance, legal or customer data carries a different level of risk from one that summarizes a public FAQ.

The Irish NCSC assessment names governance and usage policies for AI, including tools already used informally, and identifying key assets at risk as immediate priorities for public sector bodies. The same logic applies in any organization: know which assets matter, then see which agents touch them.

Argus identifies and ranks risks such as unrestricted sharing and sensitive connections, so the first ticket your team opens is for the agent that combines a wide audience with a sensitive source, not the one that happened to be created last. For a CISO, that is the difference between a dashboard of counts and a prioritized work queue. You can see the checks in more detail on the page covering Copilot Studio agent risk.

Step 5: Fix it, then prove it was fixed

Most programs stop at the ticket. Someone is asked to restrict sharing or remove a connection, they mark it done, and the record shows a closed item. Whether the agent actually changed is a separate question, and it is the one an auditor or an incident reviewer will ask.

Remediation therefore has four parts:

  1. Assign. The finding goes to the agent's owner, with the risk and the suggested action.
  2. Track. It lives in the system your teams already use. Argus supports actionable integration with tools like Jira or ServiceNow.
  3. Verify. After the owner says it is done, check the agent's real state: sharing scope, connections, permissions.
  4. Keep the evidence. Store what the check found, with the date.

Verification is where Argus differs from tools that only identify vulnerabilities. It verifies fixes and provides evidence of resolution, so a closed item is backed by what the next audit actually saw. That evidence is useful in a regulated setting, where "we asked them to fix it" carries less weight than "we checked, and here is the result."

For emergencies, ANSSI also asks for an emergency stop that revokes the agent's access and can be triggered by monitoring when processing looks abnormal. Real-time alerts on risky changes, which Argus provides, give the team the signal that makes a quick decision possible. If a sharing setting is widened on an agent with sensitive connections, you want to hear about it that day.

Step 6: Monitor changes and keep the logs

Agents change after launch. Someone adds a connector, widens the audience or hands the agent to a colleague. A control that checks only at creation misses all of it.

Two habits keep the picture accurate:

  • Daily checks of state. The configuration you approved last month is not necessarily the configuration running today.
  • Protected logs. ANSSI asks that logs be protected from modification by the agent, so the whole sequence of request, authorization, action and result can be reconstructed after an incident.

The UK NCSC likewise stresses monitoring for unusual activity across tools, workflows and connected systems. For a business manager, this is the answer to "how would we know?": a daily audit and an alert on risky changes, with a named owner who gets the finding.

What this means for each role

CISOs get a proactive position: a live inventory, a ranked list and verification evidence they can show the board. IT security teams get findings that arrive with an owner and a ticket, not a spreadsheet to reconcile. Compliance officers get the accountability chain: agent, creator, current owner, permissions, fix and proof. Business managers get a way to let teams keep building agents without an unseen build-up of risk.

The common thread is that nobody has to guess. The question "what can this agent reach?" has an answer dated today.

Where Argus fits

Argus is a security platform built for Microsoft 365. It audits all AI agents daily, records their creators, connection points and permissions, ranks risks, alerts on changes in real time and verifies that fixes were carried out, with evidence. It links each agent to its creator and current owner, and it does this without making changes to your tenant or requiring additional Microsoft licenses. Pricing is not published; the Argus page explains how the daily watch works, and the customer stories show it in Microsoft 365 environments.

Secureai

If you want to see which agents in your tenant are shared widely or connected to sensitive data, start from the SecureAi home page.

A 30-day plan

WeekActionOutput
1Inventory all agents with creator, owner, connections and sharing scopeA complete register, refreshed daily
2Rank findings; assign owners to every agent, including orphansPrioritized queue with named owners
3Remove unnecessary connections and narrow sharing on top-ranked agentsTickets raised in your service tool
4Verify the fixes and file the evidence; set alerts for risky changesAudit-ready proof of resolution

After the first month, the work becomes a routine: new agents appear in the daily audit, risky changes raise alerts, and the annual review that ANSSI and others expect becomes a confirmation, not a discovery exercise.

FAQ

What is AI agent access control?

It is the set of rules and checks that decide which identity an agent runs under, which data and tools it can reach, what it may do there and who is accountable for it. In Microsoft 365 that includes sharing scope, connections and the credentials behind each agent. It also includes the evidence that those settings are what you intended.

How do I manage AI agent permissions across a whole tenant?

Start with a full inventory that links each agent to its creator and owner. Then apply least privilege, rank the findings by risk and route them to owners through your ticketing tool. Because agents change, repeat the audit daily and verify each fix. Argus automates the audit, ranking and verification for Microsoft 365.

Who should own an AI agent?

One named person who can approve its access, monitor it and decide to stop it, as the UK NCSC's guidance describes. Ownership should move when that person changes role or leaves. Recording both the creator and the current owner makes orphaned agents visible.

How often should agent permissions be reviewed?

ANSSI asks for the agent's mission and accesses to be re-documented at least annually and after any significant modification. Agents can change much more often than that, so a daily check of their actual state keeps the annual review accurate.

Does an agent need its own permissions or the user's?

Where the agent acts for a person, it is safest when it reaches only what that person's account allows. In Copilot Studio, Microsoft lets administrators control maker-provided credentials so end users are not given the maker's reach. Auditing which agents still use maker credentials is a sensible first check.

Sources

Created with help from BabyLoveGrowth