← Back to blog

Microsoft 365 Incident Response: Cover AI Agents Too

November 18, 2026
Microsoft 365 Incident Response: Cover AI Agents Too

A Microsoft 365 incident response plan works when it answers four questions fast: what happened, who and what was involved, what can still reach our data, and has the fix held. Most plans answer those for users, mailboxes and devices. They rarely answer them for AI agents, which now sit in the same tenant with their own connections and permissions. If you cannot list your agents, name who built each one and see what each can reach before an incident, the first hours will go on discovery instead of containment.

This guide walks through the stages of M365 incident response, the evidence you must keep, the Irish and EU reporting clocks, and where AI agents fit in.

Secureai
secureai.ie
Bring AI Agents Into Incident Response
Argus audits Microsoft 365 AI agents daily, showing their creators, connections and permissions, and verifies fixes with evidence.
Visit

TL;DR

  • Microsoft defines incident response as six steps: triage, investigate, contain and eradicate, recover, resolve, then review and improve. Run all six on agent incidents as well.
  • Evidence comes before remediation. Export sign-in and audit data before you change anything, because many fixes reduce what you can see afterwards.
  • In Ireland, a personal data breach that risks individuals must reach the Data Protection Commission within 72 hours. NIS2 adds a separate 24-hour early warning for in-scope entities.
  • AI agents are the blind spot: you need a current inventory of agents, creators, connections and permissions before the incident, and proof afterwards that fixes were applied.
  • Argus audits Microsoft 365 AI agents daily and verifies remediation with evidence.

Table of Contents

The stages of Microsoft 365 incident response

Microsoft's Defender XDR documentation sets out the workflow as triage and prioritize, investigate and analyze, contain and eradicate, recover affected resources, resolve or close, and review and improve. That is the sequence most M365 teams follow in the Defender portal, and it maps cleanly onto the classic detect, contain, recover cycle.

Ireland's National Cyber Security Centre (NCSC) describes a similar plan in its 12 Steps to Cyber Security. It should cover how an incident is identified, how it is categorized and classified, how it is contained, how the root cause is investigated and how the organization recovers. The NCSC adds that the plan should be:

  • documented and tested at least annually;
  • followed by a formal, trained team;
  • inclusive of relevant stakeholders and the third parties you rely on;
  • explicit about evidence gathering, law-enforcement reporting and notification obligations.

The five-step version people search for is simply this: detect, categorize, contain, find the root cause, recover. The six-step Microsoft version adds a formal closure and a review so the lessons feed back into the plan.

What each stage asks of you

Triage. Decide how serious this is and who needs to know. A single suspicious sign-in is not the same as a confirmed compromise with data access. Write severity levels down in advance so nobody argues about them at 2 a.m.

Investigate. Establish scope: which identities, which apps, which data, which time window.

Contain and eradicate. Stop the attacker or the exposure, then remove every way back in.

Recover. Restore access and services, with tighter controls than before.

Resolve and review. Close the incident only when you can show the cause is gone, then change the plan.

The weak stages in most tenants are the second and the last: scoping, because the data on agents is missing, and verification, because nobody can prove a fix stuck.

Why AI agents change the first hour

An AI agent in Microsoft 365, such as one built in Copilot Studio, is not just a chat window. It can hold connections to SharePoint sites, mailboxes, third-party services and tools, and it can act on behalf of users. It can also be shared widely by the person who built it.

That creates three problems in an incident:

  1. Unknown inventory. If a business user built an agent last spring, security may not know it exists. You cannot triage something you cannot see.
  2. Unclear ownership. When an agent misbehaves, someone has to decide whether to disable it. Without a named creator and current owner, that decision stalls.
  3. Unclear reach. Agents that connect to sensitive sources, or that are shared with everyone, turn a small mistake into a data exposure.

Microsoft's own guidance reflects this. Its Defender XDR documentation says Defender continuously monitors Agent 365-managed agents and detects jailbreak attempts, indirect prompt injection, secret and credential leakage and suspicious access, with near-real-time alerts and Advanced Hunting queries to trace tool invocations. Microsoft's practitioner playbook for AI investigations goes further, offering queries to inventory agents, flag agents with tool configurations, and find agents with no user authentication. An agent with no user authentication matters because its actions cannot be reliably attributed to a person.

The same playbook notes that these are detection signals, not confirmed attacks, and that sustained or escalating patterns should go to the incident response team. It also stresses that AI incidents span security, IT, data governance and compliance, so the right teams should be engaged early.

Build the agent inventory before you need it

Preparation is where an agent incident is won or lost. Before anything happens, your team should be able to answer, for every agent in the tenant:

  • Who created it, and who owns it now?
  • Which data sources and services is it connected to?
  • Who can use it, and is it shared without restriction?
  • Does it touch sensitive information?
  • When did any of this last change?

A one-off spreadsheet goes stale within weeks, because agents are created and reshared constantly. This is the gap Argus is built for. It audits every AI agent in your Microsoft 365 environment daily and records its creator, connection points and permissions. It links each agent to its creator and current owner, so an on-call analyst knows who to call. It highlights risks such as unrestricted sharing and sensitive connections, ranks them by severity, and alerts in real time on risky changes. What Argus checks is laid out on the solutions page for Copilot Studio agents.

For compliance officers, the owner link is the accountability record. When a regulator or auditor asks who approved an agent's access to a data source, you have a name and a history rather than a guess.

Detect and investigate: scope, context, signal

Good investigation starts with logs, and logs have to be switched on and retained before the incident. The NCSC's Office 365 Secure Configuration Framework tells Irish organizations to enable audit logging in the Microsoft Purview compliance portal. User and admin activity is retained for 90 days, or up to one year, depending on licence. The NCSC also advises setting retention to match your auditing responsibilities. Check what your own licences give you today, because an investigation that needs data from five months ago is useless if you only keep three.

For the method, Microsoft's AI Red Team describes a scope, context, signal approach:

  1. Scope. Identify who interacted with AI systems, when, and through which services.
  2. Context. Expand to the resources that were accessed and the data that may have been exposed.
  3. Signal. Only then weigh detection signals such as prompt injection, anomalous usage or credential exposure within that chain of activity.

The point is to build a coherent account of activity rather than react to one alert. For agents, scope and context are exactly what a daily audit gives you: the agent, its creator, its connections and its permissions are already written down, so the first step is a lookup rather than a hunt.

For identity-based incidents, which remain the most common M365 starting point, Microsoft's phishing investigation playbook gives a five-stage flow. Confirm the message, scope the affected users, assess what users did and whether credentials were exposed, look for downstream activity such as suspicious sign-ins, OAuth consent grants and token abuse, then contain. Note the OAuth consent step. A consent grant to a malicious app and a risky agent connection look similar to an attacker: both are delegated access that survives a password change.

A short checklist for the first hour

  • Record the trigger and the exact time you learned of it. This also starts the reporting clock.
  • Pull sign-in and audit records for the affected identities over a generous window.
  • Check for new MFA methods, role assignments, app consents and mailbox rules.
  • List any agents the affected users created, own or can use, and what those agents connect to.
  • Open a communication log and timestamp every decision.

Contain without destroying evidence

Containment is where teams make irreversible mistakes. The rule is simple: capture first, change second. Disabling an account, deleting a rule or removing an app consent can erase the very thing that shows how the attacker got in. Export the sign-in records, the audit search results and the details of any rules or grants before you remove them, and keep a note of the search IDs so someone else can reproduce your results later. Your forensic team, insurers and regulators will work from that export folder.

For a compromised account, the order matters too. A password reset on its own does not end an attacker's access if they hold a valid session token or have added their own persistence. A sound sequence is:

  1. Preserve the minimum necessary evidence.
  2. Revoke active sessions and refresh tokens.
  3. Remove attacker-added MFA methods, registered devices, OAuth consents and mailbox or forwarding rules.
  4. Reset credentials.
  5. Isolate any suspect endpoint.

For an agent, containment follows the same logic. Identify the agent, find its owner, and decide quickly whether to restrict sharing, cut a risky connection or disable it. Microsoft's guidance for AI incidents describes staged remediation: immediate containment in the first hour, expanded remediation across related variants in the first 24 hours, and source fixes over days or weeks. The first stage includes restricting access. Because the owner is already recorded and the risk already ranked, that decision does not wait on a search for who built the agent.

The same guidance points out that AI behavior is non-deterministic, so one clean test is not proof of a fix. Each stage needs a watch period afterwards. It also says AI severity should be weighted by the deployment domain, the affected population and the nature of the content, not just the number of records.

Recover and prove the fix

Most incident reports end with a line like "remediation completed." Few can show it. For a regulator, an auditor or your board, the question after recovery is: how do you know the agent no longer has that connection, and that nobody re-enabled it last Tuesday?

This is where continuous auditing earns its place in the process. Argus verifies fixes and provides evidence that remediation was actually implemented, rather than relying on a ticket being closed. Because it audits daily and alerts in real time on changes, a risky setting that creeps back shows up as a new finding. It also works with tools such as Jira or ServiceNow, so findings turn into tracked actions with owners rather than a report nobody reads. Argus makes no changes to your tenant and does not need additional Microsoft licences, so adding it does not alter the environment you are investigating.

A closure checklist worth adopting:

  • Root cause identified and written down.
  • Every persistence mechanism removed: sessions, consents, rules, MFA methods, agent connections.
  • Evidence that each fix is in place, dated.
  • A watch period completed with no recurrence.
  • Lessons from the review turned into plan changes, with owners and dates.

Irish and EU reporting: two clocks

Incident response in Ireland has a legal layer, and it often runs on two clocks at once.

GDPR and the Data Protection Act 2018. The Data Protection Commission states that breaches presenting a risk to individuals must be reported within 72 hours of awareness. Where the breach is likely to result in a high risk, affected individuals must also be told without undue delay. Notifications go through the DPC's Breach Notification Form. Section 86 of the Data Protection Act 2018 sets the same duty in Irish law: notify without undue delay and, where feasible, within 72 hours, unless the breach is unlikely to risk data subjects. A late notice must give the reason, and you must keep a written record of every breach, including the ones you do not report.

NIS2. For in-scope entities, NIS2 requires significant incidents to be reported to Ireland's NCSC, which acts as the national CSIRT (computer security incident response team). The sequence is an early warning within 24 hours, an initial assessment within 72 hours, and a detailed report within one month. That source noted in February 2025 that the National Cyber Security Bill to transpose NIS2 was still being prepared, so confirm the current legal status with the NCSC before you rely on it.

The two regimes trigger differently. NIS2's threshold turns on operational impact, such as service disruption and effects on other entities, rather than on the type of data. As Ferraz Whitmore observes, an M365 incident can therefore be notifiable to the NCSC without involving personal data, while a personal data breach may be notifiable to the DPC without reaching the NIS2 threshold. The same analysis reports that the EU is considering changes, including a single entry point for GDPR, NIS2 and DORA reporting and a longer GDPR deadline. Treat those as proposals and check the current position when you update your plan.

GDPR and NIS2 reporting timelines compared

What this means for agents: if an agent with access to personal data was shared too widely or connected to a sensitive source, you need to establish quickly whether personal data was actually exposed. That is a scoping question, and it is far faster when you already hold a record of the agent's connections and permissions. The written breach record the Act demands is also easier to produce when your evidence is dated and kept.

Putting an Irish playbook together

A practical plan for an organization using Microsoft 365 has these parts:

ElementWhat to defineWhy it matters
Severity levelsWritten criteria from informational to confirmed widespread compromiseFaster triage, fewer arguments
RolesIncident lead, communications, legal or DPO, business ownerDecisions have a named owner
Log readinessAudit logging on, retention matched to your dutiesYou can investigate past events
Agent inventoryCreator, owner, connections, permissions, sharingScope is a lookup, not a hunt
Evidence routineExport before remediating, keep search IDsForensics and regulators can reproduce your findings
ReportingDPC 72 hours, NIS2 24 and 72 hours where in scopeYou meet both clocks
VerificationDated proof each fix is in placeClosure is defensible
ReviewLessons learned, annual test as the NCSC advisesThe plan improves

Run a tabletop exercise that includes an agent scenario: a business user's agent, shared with the whole company, connected to a finance library. Ask who finds it, who owns it, who decides to restrict it, and what evidence you would show an auditor afterwards. The gaps appear within the first ten minutes.

How Argus fits an incident response program

Argus is the agent-side layer of that program. It audits all AI agents in your Microsoft 365 environment every day and records their creators, connection points and permissions. It ranks risks such as unrestricted sharing and sensitive connections by severity and links each agent to its owner. It alerts in real time on risky changes, verifies that fixes were implemented, and feeds actions into tools such as Jira or ServiceNow.

For a CISO, that means a standing answer to the scope question. For a compliance officer, it means accountability on record. For the IT team, it means fewer surprises when an agent turns up in an investigation. You can see how it works on the Argus page, read customer stories, or start from the SecureAi home page.

Secureai

Agents will keep multiplying in your tenant. Teams that can show what exists, who owns it and that it was fixed will handle the next incident in a fraction of the time.

FAQ

What are the five steps of incident response?

The NCSC's guidance describes a plan covering detection, categorization and classification, containment, root-cause investigation and recovery. Microsoft's Defender workflow expands this to six steps by adding formal resolution and a review. Both are the same idea: find it, size it, stop it, understand it, restore and learn.

How long do we have to report an M365 breach in Ireland?

If personal data is involved and the breach risks individuals, you must notify the Data Protection Commission within 72 hours of becoming aware. If you are an in-scope NIS2 entity and the incident is significant, an early warning to the NCSC is due within 24 hours. These are separate duties, so assess both from the start.

Should we capture evidence before disabling an account?

Yes. Export sign-in records, audit search results and the details of rules, consents and connections first, then contain. Remediation can remove the data that shows how the attacker got in, and regulators and forensic teams will work from what you saved.

How do AI agents change Microsoft 365 incident response?

Agents add identities with connections and permissions that many teams have not inventoried. You need to know who built each agent, what it reaches and whether it is shared too widely. A daily audit such as Argus gives you that record in advance and evidence that fixes held afterwards.

Sources