AI Incident Response: One Alert, Handled Two Ways
Skip to content

AI in Incident Response

AI incident response is the use of AI agents and automation to carry out parts of a security incident response. Agents gather evidence, connect events, draft findings, and prepare or perform containment actions, while people make the decisions with business consequences.

The clearest way to see what changes is to follow one alert handled both ways. This guide does that, then goes through response tasks one by one and says which to automate and which to keep with people.

Why Incident Response Turned to AI

Security operations have a problem of arithmetic. Monitoring tools produce more alerts than people can investigate. Each alert, taken seriously, needs the same set of lookups: who is this user, is this address known, what else did this device do, has this happened before. A thorough investigation of one alert can take an analyst half an hour or more, and a mid-sized organization can generate hundreds of alerts a day.

The usual result is triage by neglect. Analysts investigate the alerts that look worst and close the rest on a glance. Attackers benefit, because early signs of an intrusion rarely look like the worst alert in the queue.

Earlier answers went part of the way. Better tuning reduced the number of alerts and also suppressed some that mattered. Fixed automation, where a tool runs a scripted sequence for each alert type, removed repetitive lookups but could not go beyond its script.

AI agents address the part that remained: the investigation itself. An agent reads each result and continues until it has enough to reach a conclusion. That is the work analysts never had time to do for every alert.

The effect on incident response is at the start of the timeline. The stages that cost the most time in a real incident are noticing, understanding, and scoping. Those are the stages agents accelerate.

One Alert, Handled Two Ways

At 02:10 an identity alert fires. A finance user’s session, opened from the office at 01:58, has appeared on a second network with a different browser. The sequence below compares an on-call analyst working by hand with an AI agent that investigates and holds actions for approval.

TimeAnalyst working by handAI agent with analyst approval
02:10The alert joins a queue. The on-call analyst is paged.The agent starts investigating as the alert arrives.
02:11The alert is still in the queue.The agent pulls 30 days of sign-ins for the user, the device history, and recent audit events.
02:12Waiting.It finds a forwarding address set on the mailbox at 02:04 and an application consent at 02:06, both from the second network. It checks the other users whose sessions appeared on that network.
02:13Waiting.The agent writes a case: a verdict of likely session theft, the evidence, the timeline, and proposed actions in order. Block sign-in, reset the password, revoke sessions, then remove the forwarding address and the application grant. The analyst is paged with the case.
02:20The analyst is at a laptop and opens the alert.The analyst reads the case, checks the evidence, and approves. Sign-in is blocked, the password is reset, sessions are revoked, and the forwarding address and grant are recorded and removed.
02:35The analyst has pulled sign-in history and is reading it.The agent has listed the mail the session opened and found two colleagues who received the same phishing message.
02:50The analyst finds the forwarding address and the application consent.The analyst decides to reset the two colleagues’ passwords and wake the finance manager.
03:05The analyst blocks sign-in, resets the password, revokes sessions, and removes the forwarding address and the grant.Containment has been in place for 45 minutes. The agent confirms that nothing from the second network has succeeded since.

The analyst did nothing wrong in the left-hand column and did the same work in the same order. The difference is who did the gathering and when the person’s judgment was applied. Containment came 45 minutes earlier, and the analyst spent their time deciding instead of collecting.

The agent did not decide to wake the finance manager, and it did not decide what to tell the two colleagues. Those stayed with a person. The containment itself follows the order set out in the account takeover guide: block, reset, revoke, remove, confirm.

How an Investigation Agent Works

An agent is given a goal, such as deciding whether a sign-in is malicious, and a set of tools it is allowed to use. Each tool is a defined action: search sign-in logs for a user, look up an address, list processes on a host, fetch the audit trail for a mailbox. The agent then works in a loop.

1
PlanThe agent decides what would be most useful to know next.
2
ActIt calls one of its tools with specific parameters.
3
ObserveIt reads what came back.
4
ReasonIt updates its view of what is happening.
5
Repeat or concludeIt continues until it can state a conclusion with evidence, or until it reaches a limit and hands over to a person.

The design of the investigation side determines whether the output can be trusted.

Every claim points to a record

A well-built agent cites the log entries behind each statement in its case. A conclusion without evidence should be treated as unverified.

Tools are read-only by default

Investigation tools retrieve data and change nothing. Actions that alter systems are separate, fewer, and governed by the levels of automation described later.

Limits are set

A cap on steps, time, and cost per investigation stops an agent from wandering.

Memory is controlled

Knowledge about the environment, such as which accounts are administrators or which behavior has been approved before, is supplied deliberately and can be reviewed.

The underlying language model supplies reasoning. The tools supply facts. An agent is only as good as the data its tools can reach, and that is the first thing to check in any deployment.

Response Tasks, One by One

TaskWho should do itWhy
Pulling logs and history for an account or hostAgentRepetitive, time-sensitive, and changes nothing.
Adding context: asset owner, reputation of an address, normal behavior for the userAgentA lookup the agent does the same way every time.
Linking related events across sources into one caseAgentFaster and more thorough than a person switching between tools.
Building a timelineAgentMechanical once the evidence is collected.
Searching for the same traces across all users and systemsAgentScale is where people run out of time.
Revoking sessions, isolating a standard laptopAgent, with limitsLow impact and quick to undo. Speed matters.
Isolating a server, disabling a privileged or executive accountAgent prepares, person approvesThe wrong call disrupts the business.
Cutting a supplier’s access, taking a service offlinePersonDepends on business context the agent does not hold.
Judging that scope is complete and recovery is safePersonA judgment the organization has to own.
Notification, legal steps, public statementsPersonLegal and reputational consequences.
Drafting updates and the post-incident reportAgent drafts, person reviewsThe record is already in the case.

Levels of Automation

Automation in response is set per action. Each action type can sit at a different level, and the level can rise as the record supports it.

LevelWhat happensSuits
Co-pilotThe agent investigates and recommends. A person decides and acts.New deployments and high-impact actions.
Wait for approvalThe agent prepares the action and holds it. A person approves with one step.Actions that are usually right, with occasional exceptions.
AutonomousThe agent acts first and reports afterward, within agreed limits.Low-impact actions that are quick to undo, where minutes matter.

Set the level per action type and per part of the organization. Revoking a session for a standard user might be autonomous. Isolating a production server might wait for approval. Disabling the chief executive’s account might stay at co-pilot.

In the example, containment for a finance account was at wait for approval. At autonomous, it would have happened at 02:13 and the analyst would have been told afterward.

How an Agent Differs From a Playbook Tool

Automation in security operations is not new. Orchestration tools, known as security orchestration, automation, and response (SOAR), have run fixed playbooks for years: when this alert fires, run these steps in this order.

A fixed playbook does exactly what was written. It works well for predictable tasks and stops being useful when an incident does not match the script.

An AI agent chooses its next step from what it has found. In the example, the agent checked other users whose sessions appeared on the same network because the first evidence suggested a stolen session. Nobody wrote that branch in advance. This is what lets an agent handle incidents that fall between playbooks.

Fixed automation suits tasks that should never vary, and agents suit investigation, where the right next question depends on the last answer.

What Changes for the Security Team

Introducing agents changes the work people do more than it changes how many people are needed.

Analysts review cases instead of building them. The job shifts from gathering evidence to judging it. This is more demanding in one way: an analyst has to be able to tell when a well-written case is wrong.

Decisions concentrate. With the routine investigation done, what reaches a person is the set of cases that need a human call. The proportion of an analyst’s day spent on judgment rises.

New responsibilities appear. Someone has to decide which actions may be automated, review the agent’s work, investigate its mistakes, and maintain the context it relies on. In mature teams this becomes a defined role.

Experience matters more. Junior analysts traditionally learned by doing the repetitive work that agents now take. Teams need a deliberate way to build judgment in new staff, such as reviewing agent cases alongside a senior analyst.

Coverage improves at night and on weekends. An agent investigates the 3 a.m. alert as thoroughly as the 3 p.m. one. The person on call receives a finished case instead of a raw alert, which changes what on-call work is.

The measures that show whether it is working are the time from alert to verdict, the time from alert to containment, the share of alerts investigated in full, and how often people override the agent. The number of alerts closed is a poor guide.

Where Else Agents Help

Alert investigation is the most visible use. The same approach applies to other parts of security operations that are limited by analyst time.

Threat hunting

A hunt starts from a question, such as whether any account has consented to an unfamiliar application in the past 90 days. An agent can run the searches across months of retained data, follow up on what it finds, and write up the result, including when it finds nothing.

Retrospective search

When a new threat is disclosed, an agent can check retained data for its known traces across every monitored environment within hours.

Scoping during an incident

Once one indicator is known, an agent searches every user and host for it, then for whatever those results reveal.

Exposure checks

An agent can compare a new advisory with the asset inventory and say which systems run the affected version.

Detection tuning

As an agent works alerts, it sees which rules fire on benign activity. It can propose changes, such as narrowing a rule or adding an exception, for an engineer to review.

Recurring reports

Monthly summaries and service reviews can be drafted from case records that already contain the evidence.

Controls for Automated Response

An automated action that is wrong can cause an outage. Controls keep the risk in proportion to the benefit.

Rollback

Every automated action has a documented way to undo it, and the undo is tested.

Confidence thresholds

The agent states how sure it is, acts only above a set confidence, and passes uncertain cases to a person.

Scope limits

Rules set which systems and accounts can be acted on, and how many in a given period, so one mistake cannot take down a department.

Least privilege

The agent can reach only the systems and data each task needs.

Untrusted input

Text inside logs, emails, and alerts is treated as data. An instruction hidden in a phishing email changes nothing about what the agent does.

Review

People sample automated decisions regularly, and wrong actions are treated as failures to learn from.

Evaluating an Agent Before Trusting It

No organization should give an agent authority on the strength of a demonstration. The evaluation uses the organization’s own alerts and happens before the agent can change anything.

1
Replay past incidentsRun the agent against alerts from incidents that have already been investigated by people. Compare its conclusion, its evidence, and its proposed action with what the team found.
2
Include the quiet onesTest with the early alerts from real intrusions, the ones that looked minor at the time. An agent that only handles obvious cases adds little.
3
Run in shadow modeLet the agent investigate live alerts and write cases, with no ability to act. Analysts work as usual, and the results are compared daily.
4
Measure agreement and disagreementWhere the agent and the analyst differ, find out who was right. Both kinds of error matter: threats the agent dismissed, and benign activity it would have acted on.
5
Check the evidence trailOpen a sample of cases and confirm that every statement is supported by the record it cites.
6
Test the edgesFeed it alerts with missing data, contradictory records, and text designed to mislead it, and see whether it recognizes the limits of what it knows.

Questions worth asking any supplier of an agent-based service:

  • Where is our data processed, and is it used to train models?
  • What can the agent do without a person, and how is that configured?
  • How is each automated action undone?
  • How do you detect and handle a wrong conclusion?
  • What happens when the agent is unsure?
  • Can we see the full reasoning and evidence for every case?
  • How is the agent protected from instructions hidden in the data it reads?

Limits to Plan For

Missing data

An agent cannot investigate a system that sends no telemetry, or search logs that were never kept.

Unusual but legitimate activity

A finance director signing in from a hotel abroad looks like an attack. Context the agent lacks, such as a travel booking, decides it.

New techniques

An agent may miss activity that does not resemble anything it has seen.

Over-containment

An agent tuned to act quickly will sometimes act on a false alarm. Rollback and scope limits decide how much that costs.

Major incidents

Coordination, communication, and hard decisions still need experienced people. Automation shortens the technical work and frees time for them.

Introducing It in Stages

1
Automate evidence collectionThis carries no risk to operations and gives the largest time saving.
2
Let the agent assemble casesAnalysts work from the agent’s case and report every error they find in one.
3
Run actions at wait for approvalEach proposed action is compared with what the responder decided.
4
Track overrides and rollbacksThe rate at which people overrule the agent, and the number of actions undone, form the record that later decisions rest on.
5
Move the safest actions to autonomousSession revocation for standard users is the usual first candidate.
6
Lower the level when quality dropsA change in log sources, tooling, or attacker behavior can make a trusted action unreliable.

Automation works on top of a tested process. A procedure that people have never run by hand should not be handed to an agent.

AI Incident Response With ThreatDefence

ThreatDefence combines an AI SOC platform with a 24x7 security operations center. AI agents investigate every alert by querying connected security tools, correlating endpoint, identity, network, and cloud evidence, and preparing a verdict with supporting records, the queries performed, and recommended response actions.

Operating modes are configured per customer and action type: co-pilot, wait for approval, or autonomous execution for nominated actions. Every action has a documented rollback, low-confidence cases are escalated to a person, and ThreatDefence analysts review every autonomous action.

Confirmed incidents are escalated to the SOC and DFIR team for response.

Response actions in the ThreatDefence portal: revoke sessions, reset password, isolate endpoint and block address, each with its operating mode, status and rollback

Experiencing a breach?

Our incident response team answers 24x7, with or without a retainer.

Talk to Our Experts

See how ThreatDefence automates security monitoring, compliance and cyber risk management in one platform.

    I'm an MSP, MSSP, or reseller

    By submitting this form, you accept the terms in our privacy policy.

    Frequently Asked Questions

    What is AI incident response?

    AI incident response puts an AI agent at the front of the response process. When an alert fires, the agent pulls the logs, checks history and context, links related events, and writes a case with a verdict and the evidence behind it, usually within minutes. Containment actions are proposed for approval or run automatically within limits the organization has set. People still own scope, recovery, notification, and anything with legal weight.

    Can AI fully automate incident response?

    No. Investigation and low-impact containment, such as revoking a session or isolating a standard laptop, can run without a person. Taking a server offline, cutting a supplier’s access, deciding that recovery is safe, notifying regulators, and speaking for the organization cannot. Teams that adopt agents keep their responders, whose work moves from collecting evidence to checking cases and making those calls.

    Which incident response tasks are safe to automate?

    Start with tasks that change nothing: pulling logs, adding context, linking events into one case, building the timeline, and searching for the same traces across every user and host. Next come actions that are quick to undo, such as session revocation and isolation of a standard endpoint, first with approval and later automatically. Anything that can stop a business service stays with a person.

    What is the difference between SOAR and AI incident response?

    SOAR, short for security orchestration, automation, and response, runs playbooks written in advance: a named trigger starts a fixed series of steps. An AI agent has no fixed series. It chooses each query from the result of the last, so it can follow an incident that no playbook anticipated. Most deployments use both, with the agent investigating and a playbook carrying out the approved action the same way every time.

    Is automated containment risky?

    Yes, in a specific way: a wrong action interrupts legitimate work. Blocking a finance director who really is traveling, or isolating a server during payroll, costs more than a missed low-severity alert. The risk is managed by limiting which accounts and systems an agent may act on, capping how many actions it can take in a period, requiring approval above a set impact, and testing the rollback for every action.

    What data does AI need for incident response?

    An agent needs the same sources an analyst would open: identity sign-in and audit logs, endpoint telemetry, mail and file audit records, and network and cloud logs, with enough history to compare today against the past month. It also needs context that logs do not hold, such as who owns each asset, which accounts are privileged, and what has been explained before. A source the agent cannot query is a blind spot in every case.

    How do we know an AI agent's conclusion is right?

    Every statement in the case should cite the log record behind it, so a reviewer can open the record and check. Before go-live, the agent is replayed against past incidents and run in shadow mode beside analysts, and each disagreement is traced to who was right. In production, the rates at which people override its verdicts and roll back its actions show whether it still deserves its level of authority.