Skip to main content
An old telephone switchboard with patch cords plugged into rows of jacks, drawn in dots

What is an AI SOC?

An AI SOC uses AI agents to triage, investigate, and close or escalate security alerts with evidence. How it works, how it differs from SOAR, and the risks.

Explainer
Updated 10 min read

Definition

AI SOC

Also called Agentic SOC, AI-native SOC, Autonomous SOC

An AI SOC, also called an agentic SOC, is a security operations center where AI agents do the first pass on every alert: they gather context from the SIEM, EDR, identity, cloud, and email tools, decide whether the alert is a real threat, and close it or escalate it to an analyst with the evidence attached.

How an AI SOC works

An AI SOC replaces the first human touch on an alert with an agent: a language model that can call tools, read the results, and decide what to look at next. The loop has four parts, and the large platforms and agent-native products run it differently.

  1. Trigger. An alert lands from a SIEM, EDR, identity provider, cloud service, or a user-reported email, and the agent picks it up without anyone assigning it. Where the agent runs decides which alerts qualify. Google's Triage and Investigation Agent starts five minutes after an alert by default, runs only on data ingested through Google SecOps SIEM, doesn't process alerts from the SecOps SOAR connector, and has no investigation queue, so alerts over its quota aren't analyzed automatically. Cotool takes alerts from native integrations, webhooks, email, Jira, Linear, or its API, so the source doesn't have to be a SIEM.
  2. Investigation. The agent runs queries and enrichments, reads what comes back, and plans the next step from it. Google documents its agent running searches, enriching indicators with Google Threat Intelligence, and rebuilding process trees. Microsoft's Phishing Triage Agent detonates files and URLs and runs advanced hunting across Defender data. Cotool gives each agent an isolated Linux sandbox with cloud CLIs, so it can parse a CloudTrail export or decode a payload itself, and lets it split the work: on a suspicious OAuth grant, one investigator subagent reviews 30 days of Okta sign-ins while another checks Google Drive sharing.
  3. Verdict. The agent decides whether the alert is a true or false positive and explains why. Google's agent returns a disposition, a confidence level, and a summary; Microsoft's returns a plain-language rationale and a graph of its reasoning. In Cotool, a reviewer subagent can check the verdict before it's final, and every verdict shows the evidence behind it so an analyst can reopen the investigation and ask follow-up questions.
  4. Close or escalate. False positives get resolved with the evidence on record. Real threats go to an analyst, and some platforms hand the verdict to a response playbook. When Microsoft's agent classifies an alert as a true positive, the incident stays open for an analyst. In Cotool, a response such as isolating a host in Defender or suspending a user in Okta is a state-changing tool call that waits for approval in the app or in Slack.

Analyst feedback then improves the agent. Microsoft says feedback refines its triage agent over time. Cotool turns feedback into instruction changes you review, and an AI judge grades every run against your acceptance criteria, so a broken integration such as an expired Okta token shows up the day it fails.

How it differs from SOAR and from SIEM copilots

The difference from SOAR is who writes the steps. A SOAR playbook is a fixed sequence someone built in advance for an alert type they expected; an agent chooses its next query from the last result. Google describes its approach as combining "deterministic, predefined playbooks with dynamic AI agents," and Microsoft says its triage agent applies contextual reasoning "rather than relying on static rules or predefined inputs." Playbooks still execute known response steps the same way every time, which is why many AI SOC products keep them, and teams leaving one need a migration path: Cotool imports Tines stories as agents. For the playbook model itself, see what SOAR is and where it still fits.

The difference from a copilot is who starts the work. A copilot is a chat assistant inside the SIEM: an analyst opens an alert, asks a question, and gets a summary or a query back. Elastic AI Assistant for Security provides "a chat interface where you can ask questions" about Elastic Security data, and Elastic advises users to cross-verify its advice. The alert still waits until a person opens it; an AI SOC agent starts when it arrives.

Some copilots are priced as an add-on. Microsoft Security Copilot is billed in Security Compute Units (SCUs): as of November 2026, $4 per provisioned SCU per hour and $6 per overage SCU, with 400 SCUs a month per 1,000 user licenses included for eligible Microsoft 365 E5 and E7 customers. Microsoft's Phishing Triage Agent requires Security Copilot SCU capacity, and Microsoft counts agent runs against the same SCUs as analysts' prompts in the portal, so alert volume and copilot use draw down one budget, with overage SCUs billed at the higher rate.

Why it matters for a SOC

When alerts outnumber the analysts who can investigate them, the queue forces a choice between skimming and skipping, and the result is alert fatigue. Agents change the economics: an investigation costs compute rather than analyst time, so every alert can get the same first-pass depth at 3 a.m. as at noon.

The published evidence is still thin. One of the few public studies, by Microsoft researchers, compared organizations that adopted Security Copilot with matched non-adopters and found a 30.13% reduction in mean time to resolve incidents three months after adoption, with no significant effect in the first two months; the authors caution that the design can't rule out other causes. That study measured a copilot, not autonomous triage. Measure agents on your own alerts; how to evaluate an AI SOC platform walks through a pilot.

What still needs a person

An AI SOC moves analysts from the first pass to the decisions that carry consequences.

  • Approving actions that change state. Isolating a host or disabling an account can disrupt the business. OWASP recommends human approval for high-impact actions taken by LLM agents. Cotool marks each tool action read-only or state-changing, so suspending an Okta user or isolating a host in Defender can wait for approval in the app or in Slack while lookups run on their own.
  • Working real incidents. Scoping, containment trade-offs, communication, and legal calls stay with people. Agents shorten the time to get there.
  • Defining what good looks like. Someone has to write the escalation criteria, decide which data agents may read, and review closed alerts for confident mistakes. In Cotool, that work lives in Git: agents and skills sync as YAML, so a change to escalation criteria goes through pull-request review and can be rolled back.
  • Owning the verdict. Vendors say this directly: Dropzone AI's homepage says "Your analysts own every verdict."

Kinds of AI SOC products

AI SOC products fall into four groups, depending on where the agent lives.

Agents inside a SIEM or XDR platform. These work mainly on their vendor's own data, and a source the platform hasn't ingested is one the agent can't investigate. Microsoft ships Security Copilot agents in Defender, including the Phishing Triage Agent, whose coverage of identity and cloud alerts is in preview. Google Security Operations has the Triage and Investigation Agent, built on Gemini, which runs only on SIEM-ingested data and isn't FedRAMP compliant. CrowdStrike's Charlotte AI runs multi-agent investigations across Falcon, and SentinelOne's Purple AI runs agentic investigations over SentinelOne and third-party data. See built-in security AI vs AI SOC platforms.

Agent-native platforms. These sit on the tools you already run and investigate alerts from every vendor. Cotool connects to over 100 built-in integrations plus your own MCP servers and CLI tools, runs any number of agents you configure, and can also be your SIEM if you want it to. Dropzone AI investigates across the stack through 100+ integrations and describes itself as "not a SIEM or staffed service between you and your environment." Prophet Security offers agents for investigation, hunting, and detection engineering. Intezer pairs AI-planned investigations with forensic analysis of endpoint artifacts, and 7AI assembles investigations before an analyst opens the case.

Automation platforms with agents. Teams that already run playbooks often start here. Torq combines AI agents with its hyperautomation engine, and Tines positions itself as an environment for agents, apps, and automations. See what SOAR is.

Build your own. Some teams connect a general-purpose agent such as Claude Code to their tools through MCP servers: full control, and the full maintenance burden. Connecting security tools to Claude Code shows what's involved.

Risks to plan for

The risks of an AI SOC come from the same properties that make it useful: it reads untrusted data, reaches conclusions on its own, and holds credentials to your tools.

  • Hallucination. OWASP describes models producing content "that seems accurate but is fabricated." In a SOC, the dangerous version is a confident "benign" on a real intrusion. Require every verdict to cite the queries it ran and the results it relied on, and grade a sample of runs; per-run evals like Cotool's make this continuous rather than a quarterly audit.
  • Prompt injection. Agents read attacker-controlled text: phishing emails, command lines, file contents. OWASP's indirect prompt injection entry covers exactly this case. Treat the alert payload as hostile input and limit where an agent can send data; Cotool restricts each sandbox's outbound traffic per agent.
  • Permissions. OWASP's "excessive agency" entry recommends least privilege, authorization in the downstream system, and logging every action. For example, Microsoft scopes its phishing agent's identity to read only the emails tied to alerts, not every mailbox. In Cotool, each tool action is labeled read-only or state-changing, and one click makes an agent read-only.
  • Auditability. Months later you'll need to show why an alert was closed. Check that every tool call, query, and result is kept with the verdict, not just a summary.
  • Coverage gaps. "Every alert" has limits. Google documents a quota of around 10 triage-agent investigations per hour per tenant and no investigation queue, so the rest are left for manual triage. That caps a tenant at roughly 240 agent investigations a day, however many alerts it raises. Models also refuse legitimate defensive work: on Cotool's BlueBench, two Claude models refused a material share of tasks on cybersecurity grounds. See which AI models are best for security work.

How Cotool approaches the AI SOC

Most products in the first group add agents to platforms built to centralize logs and route alerts to analysts. Google's agent runs only on data ingested into Google SecOps SIEM, and Microsoft's agents draw on Security Copilot capacity billed by the hour. Cotool starts from the agent instead, and answers the gaps above directly.

It investigates alerts from any source whether or not it stores the logs, so adopting it doesn't start with a SIEM migration. You choose the model for each agent, from Anthropic, OpenAI, Google, xAI, open-weight models, or your own endpoint, and Cotool Router moves a task to another provider or an open-weight model when one refuses security work or goes down. Every run is graded, every verdict carries its evidence, and agents live in Git. If you want AI only inside a console you already use, the built-in Microsoft or Google agents are the shorter path. Cotool fits teams that want agents across the whole stack and control over the models, tools, and actions they use.

  • SOAR: security orchestration, automation, and response; deterministic playbooks for known steps. See what SOAR is.
  • SIEM: the system that collects logs and runs detections, and the most common source of alerts an AI SOC works.
  • AI SOC analyst: vendors' name for the agent that triages and investigates a single alert.
  • MCP: the Model Context Protocol, the most common way agents connect to security tools. See security MCP servers.
  • MTTR and MTTD: mean time to resolve and to detect, the metrics most AI SOC pilots report. See SOC metrics.
  • Agentic offensive framework: attacker-side tooling in which AI agents plan and run intrusion steps. See what ARTEX is.
  • Defense factory: a continuous, agent-first operation that finds, validates, and fixes security weaknesses. See what a defense factory is.

FAQ

Is an AI SOC the same as an agentic SOC?
Yes. Vendors use AI SOC, agentic SOC, AI-native SOC, and autonomous SOC for the same idea: AI agents that investigate alerts on their own and hand analysts a verdict with evidence. "Autonomous SOC" sometimes implies the agents also take response actions without approval, so check what a vendor means before comparing claims.
Does an AI SOC replace SOC analysts?
An AI SOC does not replace SOC analysts; it changes what they spend time on. Agents take the first pass on alerts and close the clear false positives, while analysts handle escalated threats, approve containment actions, set the rules agents follow, and review whether the agents' verdicts are right.
How is an AI SOC different from SOAR?
SOAR runs playbooks that a person wrote in advance, step by step, for alert types someone anticipated. An AI SOC uses agents that decide which queries to run next based on what they find, so they can attempt alerts no playbook covers. Many platforms combine the two, with an agent producing the verdict and a playbook executing the response.
What does SOC mean in cybersecurity?
In cybersecurity, SOC stands for security operations center: the team, processes, and tools that monitor an organization for threats, triage alerts, investigate incidents, and coordinate response. An AI SOC is a SOC in which AI agents do much of the alert triage and investigation that analysts used to do by hand.
Which companies sell AI SOC platforms?
AI SOC products come from three kinds of companies. Security platform vendors add agents to their own SIEM or XDR, such as Microsoft Security Copilot agents and Google Security Operations' Triage and Investigation Agent. Agent-native vendors such as Cotool, Dropzone AI, Prophet Security, Intezer, and 7AI investigate alerts across the tools you already run, and automation vendors such as Torq and Tines add agents to workflow platforms.

Sources

  1. Microsoft Security Copilot Phishing Triage Agent in Microsoft Defender · accessed November 16, 2026
  2. Application card for Microsoft Security Copilot agents · accessed November 16, 2026
  3. Microsoft Security Copilot pricing · accessed November 16, 2026
  4. Microsoft Security Copilot Security Compute Units and capacity · accessed November 16, 2026
  5. Triage alerts with the Triage Agent - Google Security Operations · accessed November 16, 2026
  6. Respond to alerts and cases - Google Security Operations · accessed November 16, 2026
  7. Elastic AI Assistant for Security · accessed November 16, 2026
  8. Charlotte AI: Agentic Analyst for Cybersecurity - CrowdStrike · accessed November 16, 2026
  9. Purple AI: Agentic AI Security Analyst - SentinelOne · accessed November 16, 2026
  10. Dropzone AI · accessed November 16, 2026
  11. Prophet Security · accessed November 16, 2026
  12. Intezer · accessed November 16, 2026
  13. 7AI · accessed November 16, 2026
  14. Torq · accessed November 16, 2026
  15. Tines · accessed November 16, 2026
  16. Generative AI and Security Operations Center Productivity: Evidence from Live Operations · accessed November 16, 2026
  17. LLM01:2025 Prompt Injection - OWASP Gen AI Security Project · accessed November 16, 2026
  18. LLM06:2025 Excessive Agency - OWASP Gen AI Security Project · accessed November 16, 2026
  19. LLM09:2025 Misinformation - OWASP Gen AI Security Project · accessed November 16, 2026
  20. BlueBench Capability Leaderboard - Cotool · accessed November 16, 2026
  21. AI SOC Without the Black Box - Cotool · accessed November 16, 2026
  22. Introducing Cotool Router - Cotool · accessed November 16, 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