Skip to main content
A brass ship's engine order telegraph with its round dial and lever handle, drawn in dots

Incident response automation: what to run on its own and what to gate

Incident response automation runs enrichment, scoping, and containment for you. Which actions to run on every alert, which to gate, and how to verify each.

Guide
Updated 9 min read

Summary

Incident response automation uses software to run the response steps analysts used to do by hand: enriching an alert, scoping affected accounts and hosts, opening tickets, notifying owners, and containing the threat. Run every read-only step on every alert, put each action that changes another system behind an approval that names its target, and check that the action took effect. SOAR playbooks and AI agents can both run these actions; they differ in who decides the investigation steps.

Key takeaways

  1. Before automating a response action, write down what it changes, how you undo it, and the side effects its vendor documents. Reads and notifications run on every alert; containment waits for approval.
  2. Grant each responder only the actions its alert type needs, by full action name, so an identity responder can revoke sessions but can't isolate hosts.
  3. Make every approval request name the action, the target, and the evidence, and decide before the first incident what happens when nobody answers.
  4. Treat a successful API call as a request, not containment. Check sign-in logs after a session revocation and device status after an isolation.
  5. Log every tool call with its parameters and send that log to your SIEM, so automated actions get the same review as an administrator's.

The short version

Automating incident response is mostly a sorting job. Steps that only read data, such as pulling a user's sign-in history, looking up a file hash, or listing an OAuth app's scopes, can run on every alert the moment it lands, because a wrong result costs a few API calls. Steps that write to another system, such as isolating a laptop, revoking a user's sessions, or quarantining an IAM user, change what people can do at work, and several carry side effects their vendors document but few runbooks mention. Those run behind an approval until you trust the responder on that alert type.

The pressure to automate past enrichment is real. IBM's Cost of a Data Breach Report 2026 puts the global average breach at $4.99 million and says organizations that use security AI and automation extensively saved $1.93 million compared with those using none. That figure covers automation broadly, not any single action. NIST SP 800-61 Rev. 3, published in April 2025, places incident response inside cybersecurity risk management as described by the NIST Cybersecurity Framework 2.0, which is a useful reminder that a response action is a change to production, owned like any other.

Know what each action does before you automate it

Response automation goes wrong when an action does something slightly different from what its author assumed. Four common actions, from their vendors' documentation:

ActionWhat it doesDocumented limits and side effects
Isolate device (Defender for Endpoint)Cuts network access; Defender stays connectedLifts after seven days; Hyper-V host isolation cuts its VMs
revokeSignInSessions (Microsoft Graph)Invalidates refresh tokens and browser session cookiesCan lag a few minutes; skips external users
AWSCompromisedKeyQuarantineV3 (AWS IAM)Denies actions such as ec2:RunInstancesAWS-managed; deny list last edited November 9, 2026
Splunk SOAR playbook runRuns app actions in sequenceRestart cancels the run; earlier changes aren't rolled back

Defender for Endpoint's isolation has the longest list of conditions. Devices behind web proxies configured with PAC, WPAD, or static proxy settings might not recover from isolation, and Microsoft recommends selective isolation there. Devices on a full VPN tunnel can't reach the Defender cloud service once isolated. If a device is offline when isolation is sent, Defender retries for up to three days and then stops, so an administrator has to reissue it. Microsoft's own automatic isolation, part of automatic attack disruption, works only on end-user workstations, which is a reasonable line for your own automation too.

Revoking sessions in Microsoft Entra ID through Graph needs the User.RevokeSessions.All permission and returns success before revocation finishes:

POST https://graph.microsoft.com/v1.0/users/{id}/revokeSignInSessions

Microsoft notes there might be a delay of a few minutes before tokens are revoked, and the call doesn't revoke sessions for external users, who sign in through their home tenant. AWS attaches its quarantine policy to IAM users whose keys are exposed publicly and says not to remove it; you can also attach it to users, groups, and roles yourself. The policy also denies cloudtrail:LookupEvents, so an investigation can't run on the quarantined identity's credentials.

Playbooks and agents split the work differently

A SOAR playbook is a fixed sequence. In Splunk SOAR you link actions provided by apps, such as Okta's set password or enable user, in a visual editor. Tines calls the same unit a story, a set of connected actions it describes as analogous to a playbook. Approval is a block in the flow: Splunk SOAR's Prompt block sends a question to a user or role with a required response time in minutes. Two documented details matter for containment. When a prompt goes to a group, the first user to approve resumes the run. And Splunk advises against using unauthenticated external prompts to approve sensitive actions like quarantining hosts, recommending SAML or role-based access instead.

An agent chooses its investigation steps at run time, from the tools it has been granted, and the containment action stays a tool call that can require approval. What SOAR is covers where fixed playbook branches break, and AI incident response covers what agents do in each investigation phase. This page is about the actions at the end of both.

Cotool's response agents are written in natural language and triggered by webhook, API, cron, Slack, Jira, Linear, or email. The Endpoint Alert Responder shown on that page has six tools enabled: Query SIEM Logs, List Alerts, Retrieve Okta User, Take Action on Hosts, Revoke Okta User Sessions, and Create FireHydrant Incident. Four read or record, and two change state, which is the mix every responder needs to be explicit about.

Step by step

1. Inventory actions and grant them per responder

For each alert type, list the response actions, the system each one writes to, the API permission it needs, and how to undo it. Then give each responder only that list. In Cotool, agents can use only the tools you explicitly grant and can't discover others. In Cotool's agents-as-code format, toolNames lists the actions an agent may call; a bare integration id grants every action in that integration, so name full actions for any responder that can write.

2. Run reads on every alert

Enrichment, scoping, and the timeline run first and unattended. When a Cotool response agent triages an alert, it reads the alert context and timeline, investigates with its tools, adds a final comment with evidence, conclusion, and gaps, and escalates confirmed malicious alerts for human review. Starting that triage requires both the alert.manage and agent.execute permissions.

3. Gate state-changing actions with a named approval

An approval request should say which action, on which user or host, and why, with the evidence attached. Decide the timeout before you need it: in a SOAR prompt it's the required response time; in an agent it's an instruction about what to do when nobody answers. In Cotool, each tool is either read-only and automatic, state-changing and waiting for approval in the app or Slack, or state-changing and allowlisted, such as posting to a security channel. One click makes an agent read-only, which is the switch to reach for when a responder misbehaves mid-incident. For channel design and approval buttons, see running an AI SOC from Slack and Teams.

4. Write the exclusions down

List the targets the responder must never act on alone: servers, Hyper-V hosts, devices behind full-tunnel VPNs or PAC proxies, break-glass admin accounts, and external users your revocation won't reach. Put them in the playbook's conditions or the agent's instructions, and review them like code. Cotool agents and skills sync from GitHub as YAML, about every 5 minutes, with each change recorded as a version, so widening what an agent may do alone is a reviewed pull request.

5. Verify that the action took effect

A 200 response means the request was accepted. After a session revocation, query sign-in logs for new token activity once the documented delay has passed. After an isolation, check the device status, and for an offline device, set a follow-up before the three-day retry window closes. Because Defender lifts isolation after seven days, open a task to decide before it lifts on its own.

6. Record every action and measure the loop

Every automated action needs a record a person can reverse from, since a Splunk SOAR restart cancels a run without rolling back its changes. Cotool's audit log records all agent executions, tool calls and their parameters, and configuration changes, and exports through an API for SIEM ingestion. An AI judge grades every run against the acceptance criteria you set, so a responder that skips a check shows up in the scores. Track time to contain, approval wait time, and how often a person reverses an automated action; SOC metrics covers how to measure the first two.

Common mistakes

  • Treating the API response as containment. Graph session revocation can lag a few minutes, and Defender isolation on an offline device waits for the device to reconnect.
  • Granting a whole integration. A responder that only needs to read Okta users can also suspend them if you grant every Okta action.
  • Approvals anyone can click. Group prompts resume on the first approval, and Splunk warns against unauthenticated approvals for host quarantine.
  • No plan for silence. An approval with no timeout rule leaves a compromised account active, or acts with nobody watching.
  • Forgetting automatic expiry. Defender isolation lifts after seven days whether or not the investigation is finished.
  • Automating containment on servers first. Isolating a Hyper-V host takes its child virtual machines offline too. Start with workstations.
  • No change record. Without a log of what the automation changed and with which parameters, nobody can roll a bad run back.

Checklist

  • Every response action is listed with its system, permission, undo step, and documented side effects.
  • Each responder has its own grant, by full action name, and nothing more.
  • Read-only steps run on every alert without approval.
  • State-changing actions name their target and evidence in the approval request, use authenticated approvers, and have a timeout rule.
  • Servers, hypervisor hosts, break-glass accounts, and external users are excluded from unattended action.
  • A follow-up check confirms each action took effect, with reminders for retry windows and automatic expiry.
  • Tool calls and parameters are logged and exported to your SIEM.
  • Changes to what responders may do alone go through review.

FAQ

What is incident response automation?
Incident response automation is software that runs incident response steps without an analyst performing each one: enrichment, scoping, ticketing, notification, and containment actions such as isolating a device or revoking a user's sessions. It runs as SOAR playbooks, as AI agents, or as response actions built into tools like EDR.
Which incident response steps should be automated first?
Automate the read-only incident response steps first: pulling sign-in history, looking up hashes and domains, listing an app's permissions, and building a timeline. These steps change nothing in other systems, so a wrong result costs only API calls and can run on every alert.
Should containment actions be fully automated?
Containment actions should run automatically only where the blast radius is small and the undo is simple, such as a managed workstation. Isolating a server or a Hyper-V host, or revoking an executive's sessions, should wait for a named approval. Microsoft's own automatic isolation in Defender for Endpoint applies only to end-user workstations.
What is the difference between SOAR and AI agents for incident response automation?
In SOAR incident response automation, a person writes the playbook's branches in advance and the platform runs them in order. An AI agent chooses its investigation steps at run time from the tools it has been granted. In both, the containment action itself can stay a fixed, approved step.

Sources

  1. Cost of a Data Breach Report 2026 | IBM · accessed November 19, 2026
  2. SP 800-61 Rev. 3, Incident Response Recommendations and Considerations for Cybersecurity Risk Management · accessed November 19, 2026
  3. Take response actions on a device in Microsoft Defender for Endpoint · accessed November 19, 2026
  4. user: revokeSignInSessions - Microsoft Graph v1.0 · accessed November 19, 2026
  5. AWSCompromisedKeyQuarantineV3 - AWS Managed Policy · accessed November 19, 2026
  6. Use playbooks to automate analyst workflows in Splunk SOAR (Cloud) · accessed November 19, 2026
  7. Require user input using the Prompt block in your Splunk SOAR (Cloud) playbook · accessed November 19, 2026
  8. Stories | Docs | Tines · accessed November 19, 2026
  9. AI Alert Triage and Incident Response Agents | Cotool · accessed November 19, 2026
  10. AI SOC Without the Black Box | Cotool · accessed November 19, 2026
  11. Tools and Integrations - Cotool Documentation · accessed November 19, 2026
  12. Response Agents as Code - Cotool Documentation · accessed November 19, 2026
  13. Working Alerts - Cotool Documentation · accessed November 19, 2026
  14. Audit Logs - Cotool Documentation · accessed November 19, 2026

Keep reading

Attackers are scaling with tokens

Cotool helps defenders operate at machine speed. See how security teams are scaling Detection & Response.

Book a demo