← Back to blog

M365 Security Assessment: What to Check and How to Run One

November 13, 2026
M365 Security Assessment: What to Check and How to Run One

A Microsoft 365 security assessment is a structured review of your tenant’s identity, data, app, device and collaboration settings. It produces a ranked list of gaps and a plan to close them. Most assessments stop at configuration. In 2026 the larger blind spot is AI agents: who created them, what they connect to and who can use them. That is why we audit them daily with Argus, our security platform for Microsoft 365 environments, rather than leaving them to a once-a-year review.

This guide is for IT security teams, compliance officers, business managers and CISOs. It explains what to examine, how an assessment runs, how to rank findings and how to prove that fixes worked.

Secureai
Bring AI Agents Into Your Security Assessment
Argus audits Microsoft 365 AI agents daily, showing who created them, what they connect to, and whether remediation has been verified.
Visit

TL;DR

  • An assessment compares your tenant against a baseline across identities, data, apps, devices and infrastructure, then ranks the gaps.

  • A finding only counts once someone owns it, fixes it and you can show evidence that the fix is in place.

  • AI agents, including those built in Copilot Studio, need their own inventory: creator, owner, sharing scope, connections and permissions.

  • In Ireland, GDPR security duties and NCSC guidance both expect regular testing and evaluation of your measures, so keep dated records.

  • Argus audits all AI agents in Microsoft 365 every day and verifies that remediation happened, with evidence.

Table of Contents

What a Microsoft 365 security assessment is for

The purpose is simple: know where you stand, then decide what to fix first. Ireland’s National Cyber Security Centre (NCSC) puts it plainly in its Office 365 Secure Configuration Framework: understanding your current Microsoft 365 security posture is critical to identifying gaps and prioritizing a roadmap of controls that reduce cyber risk.

That framework was developed by the NCSC with Microsoft and Ekco to guide Irish Government departments in configuring Office 365 to a high level of security. You do not need to be a government body to use it as a reference. It gives you a vocabulary that auditors, boards and regulators in Ireland already recognize.

A good assessment answers four questions:

  1. What is configured today, and how does it differ from the baseline?

  2. Which differences create real risk, given what your data and users do?

  3. Who is responsible for each fix, and by when?

  4. How will you know the fix worked?

Most published assessments are strong on the first two and thin on the last two. Closing that gap is what turns a report into a lower risk.

The controls an assessment should examine

Microsoft Secure Score is the usual starting point. The NCSC framework describes it as a security analytics tool that measures your organization across five areas: identities, data, apps, devices and infrastructure. A higher score means more recommended practices are in place. It is a useful measure of configuration, and it is where the assessment begins.

AreaWhat to reviewExample checks
IdentitySign-in methods, privileged roles, risky sign-insNo legacy authentication, privileged accounts with phishing-resistant methods, just-in-time role activation
DataLabels, loss prevention, auditSensitivity labels configured, data loss prevention policies enabled, audit logging on
AppsConsent, sharing, connectionsWhich apps and agents hold permissions, who they are shared with
DevicesCompliance and access conditionsDevice-based access rules, managed device coverage
InfrastructureTenant-wide settingsExternal sharing defaults, admin separation
AI agentsInventory, ownership, reachCreator, owner, sharing scope, data sources, tools, connected agents

Identity

Microsoft’s guidance on configuring Microsoft Entra for increased security lists concrete checks. They include no legacy authentication sign-in activity, all high-risk sign-ins triaged, all risky workload identities triaged, phishing-resistant methods registered for privileged accounts, and privileged roles activated just in time rather than permanently active.

The NCSC’s 12 Steps to Cyber Security adds the management side: an effective identity and access management programme to control access to information, and a vulnerability management programme that runs from identification through to remediation.

Data

Microsoft’s guidance on configuring Microsoft Purview for increased security names checks such as sensitivity labels configured, data loss prevention policies enabled, insider risk management policies enabled for risky AI usage, communication compliance monitoring configured for Copilot, and Purview audit logging enabled.

If your organization holds a Purview license at A5, E5 or G5 level, Microsoft’s Copilot security and governance documentation describes further options. These include preventing Copilot and agents from processing certain sensitive files with data loss prevention, alerts on risky AI use such as attempted prompt injection through insider risk management, and prompt and response review in Data Security Posture Management for AI. Check your licensing before you plan around any of these.

Apps and AI agents

This is the row most assessments leave short. The same Microsoft document says its Copilot controls cover Copilot, Copilot Chat, Microsoft 365 prebuilt agents, and agents created in Copilot Studio and published to Microsoft 365 channels. Built-in protections include blocking prompt injection attacks, blocking harmful content and detecting protected material. Those protections are the platform’s. Your assessment still has to establish which agents exist in your tenant, who built them and what they can reach.

How an assessment is carried out

The mechanics vary by team, but the sequence is stable. The NCSC framework recommends a gap analysis of security and controls, followed by a remediation plan with a prioritized set of mitigation actions and an ongoing roadmap that fits your risk tolerance, skills, resources and budget.

1. Fix the scope and the baseline

Decide which tenants, business units and services are in scope, and choose the baseline. For Irish organizations the NCSC framework is a natural choice, alongside Microsoft’s own guidance for Entra and Purview. Write the baseline down. A finding is only meaningful against a stated standard.

Where it goes wrong: scope that stops at users and mailboxes. Agents, connectors and shared automations sit outside many scopes simply because nobody listed them.

2. Collect configuration evidence

Export the current state: Secure Score and its recommendations, conditional access policies, role assignments, sharing settings, label and loss prevention policies, and audit configuration. Date every export. Evidence without a date cannot show change over time.

3. Build the inventory

List identities, apps and agents. For agents, the inventory needs more than a name. You want the creator, the current owner, who the agent is shared with, what data sources and tools it declares, and what permissions it has requested and been granted.

4. Analyze the gaps

Compare each item with the baseline and mark the difference. Keep the language factual: what the setting is, what the baseline expects, what could happen as a result.

5. Rate and rank

Rank by likelihood and impact, as discussed below. Assign each finding an owner and a due date.

6. Remediate and verify

Make the change, then confirm it from the system itself, not from the ticket status. This is the step that most often gets skipped.

What you should expect to receive

A useful output has five parts: the baseline used, the inventory, the findings with severity, the owner and due date for each, and the evidence that closed findings are actually closed. If a report gives you a score and a list of recommendations but no owners and no verification, you have the first half of an assessment.

Why AI agents need their own line in the assessment

Agents behave differently from ordinary configuration. They are created by people across departments, they can be shared widely with one click, and they can connect to data sources and tools. A tenant setting reviewed in March says little about an agent built in September.

Microsoft publishes the fields that describe an agent. The AgentsInfo table in Defender XDR advanced hunting records, for each agent, its ID and name, its owners, who it is shared with (users and security groups), its permissions including requested and granted state, approval state and consent, its declared data sources and tools, any MCP servers, connected agents and lifecycle status. That list is a good checklist for what your own inventory should contain.

On the identity side, Microsoft describes how Entra and Agent 365 give visibility into agent identities, including agents with an Entra Agent ID, agents you register yourself and shadow agents. It also describes extending conditional access and identity protection policies from users to agents to enforce least privilege. The Microsoft 365 admin center’s Agent 365 overview adds central visibility and usage insights with security signals for administrators.

Why AI agents need their own line in the assessment

So the data exists. The practical question is who reads it, how often and what happens next. A quarterly look will miss an agent that was shared tenant-wide on Tuesday and connected to a sensitive site on Wednesday.

The questions to ask about every agent

  • Who created it, and who owns it now?

  • Who can use it: one person, a team or everyone?

  • What does it connect to, and is any of that sensitive?

  • What permissions has it requested, and which were granted?

  • Is it still needed, and what is its lifecycle status?

When the answers live in different consoles, the assessment becomes a manual exercise that is out of date by the time it is finished.

Ranking findings and proving the fix

The NCSC’s guidance for NIS2 entities, Members in NIS2 entities, advises a risk assessment that weighs the likelihood and impact of cyber risks, identifies the gaps, and creates a prioritized action plan implemented in order of risk. Apply that logic to findings.

A simple ranking approach

SeverityTypical exampleExpected response
HighAgent shared with everyone and connected to sensitive dataAct immediately, owner confirms within days
MediumAgent with broad permissions but limited sharingScheduled fix with an owner and date
LowUnused agent with no sensitive connectionReview at next cycle, retire if not needed

Two factors move a finding up: reach (how many people can use it) and sensitivity (what it touches). An unrestricted sharing setting on an agent connected to HR or finance content outranks a stricter one on an agent that only reads public pages.

Verification is the difference

A closed ticket is a claim. Verification is the system showing the new state. For each high and medium finding, record the before state, the change, the date and an after state taken from the source. Auditors and boards tend to ask for exactly this.

This is where Argus is built differently. It does not only identify vulnerabilities, it verifies fixes and provides evidence of resolution. It also does this without making changes to your tenant and without requiring additional Microsoft licenses.

Accountability and data protection in Ireland

For Irish organizations, the assessment is also evidence. The Data Protection Commission’s data security guidance explains that GDPR Articles 25 and 32 require data protection by design and by default and security measures appropriate to the risk. That includes a process for regularly testing, assessing and evaluating how well those measures work. The same guidance says controls built on individual and organizational accountability are an important part of protecting personal data, and that the duty applies to every controller and processor regardless of size.

If you rely on Microsoft as a cloud provider, the DPC’s guidance for engaging cloud service providers adds that a controller must be satisfied the provider’s security standards are sufficient and that a process exists for regularly evaluating the measures. Your own tenant configuration, including agents, is part of that picture.

The legal framework in Ireland is the Data Protection Acts 2018, as amended, which give further effect to the GDPR, with the Data Protection Commission as the regulator, as summarized in Legal 500’s Irish country guide. The same guide notes that the EU AI Act has had direct effect in Ireland since August 2024 and applies in phases, with most of it applying from August 2026.

The NCSC’s 2026 AI Cyber Security Risk Assessment recommends a whole-lifecycle approach to AI cyber security risk, treating security as a requirement from design to end-of-life, and points out that GDPR compliance is a critical consideration when implementing AI. It is addressed to public sector bodies, but the lifecycle idea transfers well to agents: an inventory with creators, owners and lifecycle status is the practical form of it.

If you hold or pursue certification, NSAI’s ISO/IEC 27001 page explains that it requires identified risks, controls applied through a documented risk treatment process, and processes for monitoring, internal audit and continual improvement. Certification does not by itself demonstrate legal compliance, but it can give structured evidence of controls that contribute to data protection obligations. Dated assessment records, with verified fixes, feed directly into that evidence.

What compliance officers can ask for

Linking each agent to a named creator and current owner turns an abstract policy into something you can enforce. When a regulator, auditor or board member asks who is responsible for an agent and what it can access, you should be able to answer from a record, not from memory.

Keeping the assessment current with Argus

A baseline review gives you a starting position. The agent layer changes daily, so the review has to as well. Argus audits all AI agents in your Microsoft 365 environment every day and records their creators, connection points and permissions. It identifies and ranks risks such as unrestricted sharing and sensitive connections, and it links each agent to its creator and current owner so accountability is clear.

Secureai

When something risky changes, Argus raises real-time alerts. It supports actionable integration with tools like Jira or ServiceNow, so findings reach the people who own the fix. Then it verifies that the fix was made and keeps the evidence. You can read how it works on the Argus page, see what Argus checks in Copilot Studio agents, or read customer stories from Microsoft 365 teams. To talk about your tenant, start at secureai.ie.

A short checklist to start this month

  1. Write down your baseline: the NCSC framework plus Microsoft’s Entra and Purview guidance.

  2. Export your Secure Score and its recommendations, and date them.

  3. Review privileged roles, legacy authentication and risky sign-ins.

  4. Confirm audit logging, sensitivity labels and loss prevention are on.

  5. Build an agent inventory with creator, owner, sharing scope, connections and permissions.

  6. Rank findings by reach and sensitivity, and give each an owner and a date.

  7. Verify every high and medium fix from the system and file the evidence.

  8. Repeat on a schedule that matches how fast agents change.

FAQ

What is the difference between a Microsoft 365 security assessment and Secure Score?

Secure Score is a measure inside Microsoft 365 that rates your configuration across identities, data, apps, devices and infrastructure. An assessment uses it as one input, then adds scope, a baseline, an inventory, ranked findings, owners and verification. Think of the score as a gauge and the assessment as the whole inspection.

Which Microsoft 365 security controls matter most?

Start with identity: no legacy authentication, risky sign-ins triaged, phishing-resistant methods for privileged accounts and just-in-time admin roles. Then data controls such as sensitivity labels, loss prevention and audit logging. Finally, cover agents: who built them, who they are shared with and what they connect to.

How often should we assess our Microsoft 365 security posture?

The DPC expects regular testing and evaluation of your security measures but does not set a single interval. Tenant-wide settings suit a scheduled review. Agents change faster, so a daily audit of them, as Argus performs, keeps the inventory and risk ranking current between reviews.

How do we assess AI agent security in Microsoft 365?

Inventory every agent with its creator, owner, sharing scope, data sources, tools and permissions, then rank by reach and sensitivity. Fix the highest risks first and confirm the fix from the system. Argus does this daily and links each agent to its creator and owner.

Who should own the findings?

Each finding needs one named owner, usually the agent’s creator or the team lead responsible for the service, with security tracking progress. A named owner and a due date turn a finding into a task that can be verified.

Sources