> ## Documentation Index
> Fetch the complete documentation index at: https://docs.contraforce.com/llms.txt
> Use this file to discover all available pages before exploring further.

# What Is Agentic Security Delivery? A Practical Guide

> Learn what agentic security delivery is, how AI agents investigate and respond under human governance, and how it differs from rule-based automation.

> [!NOTE]
>
> **Who is this for?** Security Leaders, CISOs, MSP/MSSP Partners

**October 2026** · ContraForce Team · 10 min read

**Agentic security delivery** is an approach to cybersecurity in which AI agents investigate, decide, and act on security work within policies your team sets, instead of relying on people to execute every step. It combines agentic AI, security operations, and governed execution, with the goal of delivering security outcomes faster, more consistently, and at scale.

The important phrase is "within policies your team sets." Agents take on the repetitive work, while people define the boundaries, approve high-impact actions, and own the outcome.

## Agentic Security Delivery Defined

Traditional security delivery is human-driven: tools raise alerts, and analysts investigate, decide, and act. Agentic security delivery hands that execution to goal-driven AI agents that reason across systems, choose a sequence of actions, and carry them out through the tools you already run.

Three ingredients define it:

* **Agentic AI**: agents that pursue an objective, such as "investigate and contain this incident," rather than answering one prompt at a time
* **Security operations**: the detection, investigation, response, and reporting work your team delivers every day
* **Governed execution**: approved actions run automatically, while high-impact actions wait for a person to approve them

If your leadership team needs a one-sentence version, use this:

> Agentic security delivery is an AI-native security operating model in which goal-driven AI agents orchestrate, execute, and continuously improve security operations, enabling faster threat response, proactive risk reduction, and scalable security outcomes under human governance.

## Agentic vs. Traditional Security Delivery

| Area | Traditional security delivery | Agentic security delivery |
| - | - | - |
| Operating model | Human-driven | AI-agent-driven with human oversight |
| Threat detection | Rules, alerts, and analyst investigation | Agents analyze and correlate signals across tools |
| Incident response | Analysts investigate and execute playbooks | Agents investigate and execute approved response actions |
| Vulnerability management | Periodic scans and manual prioritization | Continuous discovery and risk-based prioritization |
| Compliance | Manual evidence collection | Automated evidence gathering and continuous monitoring |
| Security reporting | Manually built dashboards | AI-generated insights and recommendations |
| Scalability | Grows mainly by adding security staff | Handles more volume per analyst, so headcount grows more slowly than workload |

## How Agentic Security Delivery Works

Most agentic security delivery systems have five components:

1. **Specialized AI security agents.** Each agent specializes in one or more security functions, such as threat detection and investigation, incident response, identity and access management, cloud security posture, vulnerability assessment, or compliance monitoring.
2. **An orchestration layer.** A platform coordinates agents toward a shared objective. For example, an investigation agent that confirms a compromised account hands off to a response agent that contains it, while a reporting agent documents the case.
3. **Security tool integration.** Agents act through the technology you already have: security information and event management (SIEM) and security orchestration, automation, and response (SOAR) platforms, endpoint detection and response (EDR), identity platforms, cloud security tools, vulnerability scanners, and IT service management systems.
4. **Policy-bound autonomous decisions.** Within defined policies and risk thresholds, agents perform approved actions on their own, such as investigating suspicious activity, collecting forensic evidence, opening tickets, or recommending endpoint isolation. High-impact actions should require human approval.
5. **Continuous learning and optimization.** Agents use investigation outcomes, analyst feedback, and updated threat intelligence to improve their recommendations and workflows. This doesn't necessarily mean the underlying AI model retrains itself.

## Agentic AI vs. Security Automation

This is the distinction that matters most. Security automation follows predefined rules, while agentic security works toward an objective and decides how to reach it.

| Dimension | Rule-based automation | Agentic security |
| - | - | - |
| Instruction | "If a suspicious IP address is detected, block it." | "Determine whether this IP address is a threat and respond appropriately." |
| What happens | The rule fires the same way every time | The agent investigates the IP, correlates endpoint and identity activity, weighs business impact, decides whether blocking is justified, and starts an approved response |
| Unexpected situations | Handled only if someone wrote a rule for them | The agent adapts its investigation to what it finds, within set boundaries |
| Main risk | Rigid rules miss context or misfire at scale | Incorrect decisions or unintended actions when governance is weak |

In short: **automation executes predefined instructions. Agentic AI determines which actions to take to achieve a defined security objective, within established boundaries.**

That flexibility brings new risks, including prompt injection, excessive permissions, incorrect decisions, and unintended actions. Strong governance, least-privilege access, testing, and auditability are essential, which is why governance is its own layer in the operating model below.

## Example: Agentic Response to a Suspicious Sign-In

Consider a service provider that monitors 200 customers. An impossible-travel incident arrives: a user signed in from an unfamiliar country minutes after signing in from their usual office.

**The traditional approach:**

1. The incident waits in the queue until an analyst is free, which can take hours when the queue is deep or the shift is short-staffed.
2. The analyst pivots between consoles to pull sign-in logs, device activity, and related alerts.
3. The analyst decides whether the account is compromised.
4. The analyst manually revokes sessions, resets the password, and isolates the device if needed.
5. The analyst writes notes, updates the ticket, and notifies the customer.
6. Someone compiles the activity into a report at the end of the month.

**The agentic approach** splits the same work into roles. Depending on the platform, one agent can fill several of these roles, or each role can be a separate specialized agent:

| Role | Responsibility |
| - | - |
| Triage agent | Picks up the incident as it arrives and identifies the affected users, devices, and IP addresses |
| Investigation agent | Correlates sign-in logs, audit activity, device timelines, and threat intelligence |
| Classification agent | Decides whether the incident is a True Positive, Benign Positive, or False Positive, or marks it Undetermined, by following your standard operating procedures (SOPs) |
| Response agent | Proposes containment, such as revoking sessions and resetting the password, and runs it when its confidence and your policies allow |
| Human analyst | Approves or rejects high-impact actions, such as disabling an executive's account, and notifies the customer |
| Validation agent | Confirms containment held, such as checking for new sign-ins after sessions were revoked |
| Reporting agent | Documents findings, evidence, and actions taken on the incident |

**The result:** a coordinated, largely automated workflow in which routine containment can start as soon as the incident arrives, and analysts spend their time on the cases that need judgment. Actual improvements depend on integration quality, policy controls, and operational maturity.

## Which Security Services Can Become Agentic?

Agentic delivery applies across most security service areas:

| Security domain | Agentic capabilities |
| - | - |
| Security operations center (SOC) | Autonomous alert triage, investigation, and response orchestration |
| Vulnerability management | Continuous risk assessment and remediation coordination |
| Cloud security | Misconfiguration detection and policy-based remediation |
| Identity security | Access anomaly detection and access review automation |
| Application security | Automated code security analysis and remediation recommendations |
| Governance, risk, and compliance (GRC) | Continuous control monitoring and evidence collection |
| Threat intelligence | Threat correlation, enrichment, and proactive analysis |
| Third-party risk | Continuous vendor risk monitoring and assessment |

Many organizations start with the SOC, because incident volume and response speed put the most pressure on analysts there. That's also where ContraForce focuses: incident investigation and response, such as isolating devices, disabling accounts, and revoking sessions. ContraForce doesn't patch systems, scan for vulnerabilities, harden configurations, or remediate compliance findings.

## The Agentic Security Delivery Operating Model

A typical agentic security delivery architecture has five layers, each feeding the next:

| Layer | What it includes |
| - | - |
| 1. Security data sources | Security telemetry, cloud environments, endpoints, identity systems, vulnerability scanners, and applications |
| 2. Agentic AI platform | AI models, specialized security agents, memory and context management, reasoning, workflow orchestration, and tool integrations |
| 3. Security decision and execution | Risk assessment, investigation, prioritization, remediation recommendations, and approved automated actions |
| 4. Governance and human oversight | Security policies, identity permissions, approval workflows, audit trails, AI evaluation, and human-in-the-loop controls |
| 5. Security outcomes | Reduced mean time to investigate (MTTI), reduced mean time to respond (MTTR), faster containment, audit-ready evidence, and greater operational efficiency |

Treat layer 5 as objectives rather than guaranteed benefits. You reach them by getting layers 1 through 4 right, especially governance.

## Human Oversight and Security for AI Agents

Agentic security has two dimensions, and a complete strategy addresses both:

1. **Security delivered by AI agents**: using agents to operate security services such as the SOC, vulnerability management, cloud security, and GRC
2. **Security for AI agents**: protecting the agents themselves against unauthorized actions, prompt injection, data leakage, privilege misuse, and other agent-specific risks

These controls reduce the risk of delegating work to agents:

| Control | Why it matters |
| - | - |
| Least-privilege permissions | An agent can only do what its access allows, so grant only the actions it needs |
| Approval gates | A person signs off before actions with a large blast radius, such as disabling a VIP's account |
| Confidence thresholds | Agents take response actions on their own only when their confidence meets a threshold you set; below it, the agent doesn't act and leaves the response to an analyst. Tune the threshold against observed accuracy |
| Scoped autonomy | Limit autonomy by severity, incident type, or working hours, then expand it as trust grows |
| Untrusted input handling | Incident data such as email subjects, file names, and command lines can carry attacker-written instructions (indirect prompt injection), so agents must treat it as evidence, never as commands |
| Verdict review | Spot-check incidents the agent closed as Benign Positive or False Positive, because a wrong close hides a real threat |
| Audit trails | Every agent decision and action is recorded and reviewable |
| Testing and evaluation | Have agents investigate your real incident types without taking response actions, and check their conclusions before you let them act |

## Business Value of Agentic Security Delivery

For enterprises and security service providers alike, agentic security delivery can offer four strategic advantages:

* **Operational efficiency**: Reduce repetitive security work so analysts can focus on complex threats.
* **Speed**: Accelerate investigation, prioritization, and incident response.
* **Scalability**: Support growing security workloads without growing headcount at the same rate.
* **Outcome-based delivery**: Shift from measuring activities performed to measuring risk reduction and security outcomes.

## Agentic Security Delivery for MSSPs vs. Enterprise CISOs

The goals, operating model, architecture, commercial value, and roadmap look different depending on who's adopting agentic security delivery:

| Dimension | MSP or MSSP | Enterprise CISO |
| - | - | - |
| Primary goal | Serve more customers at consistent quality without growing headcount at the same rate | Transform internal security operations and reduce risk |
| Operating model | One team and one set of procedures applied across many customers | One organization's procedures, owned by internal teams |
| Architecture priority | Multi-tenant management, per-customer policies, and separation between customers | Integration with the internal security stack and identity governance |
| Commercial value | Room for higher margin per customer and new outcome-based service tiers, net of AI consumption costs | Faster response and more incidents handled by the same team, weighed against AI consumption costs |
| Roadmap | Standardize SOPs across customers, pilot investigation-only agents on routine incident types for a few customers, then expand to more customers and to automatic response as accuracy is proven | Pilot investigation-only agents on the highest-volume, lower-risk SOC incident types, then extend automatic investigation and response as agents prove accurate on your incidents |

## How ContraForce Approaches Agentic Security Delivery

ContraForce applies this model to incident response for MSPs, MSSPs, and internal security teams. Here's how each principle maps to the platform:

| Principle | How ContraForce implements it |
| - | - |
| Agents that investigate and respond | Each Security Delivery Agent investigates and classifies incidents from connected tools such as Microsoft Sentinel, Defender XDR, and CrowdStrike, and compiles a Gamebook response when one is warranted |
| Grounded in your procedures | Classification and Response SOPs that you link to each agent |
| Approved actions through existing tools | Gamebooks run response actions such as **Isolate Endpoint**, **Invalidate Existing Sessions**, and **Reset User Password** |
| Scoped autonomy | **Mode** per severity (**Manual** or **On Queue**), plus Agent Shifts that set when an agent investigates on its own |
| Confidence thresholds | A 0–100 **Confidence to respond** setting per severity; below it, an **On Queue** run builds no Gamebook and leaves the response to an analyst |
| Approval gates | An agent's Gamebook runs without approval only when **Allow agent to run gamebooks** is on for the incident's severity and the workspace's **Allow service providers to run gamebooks** setting is on. Agents submit Gamebooks the way a service provider does, and that workspace setting is off by default. Workspace Policies can still block an action or require approval, including policies that apply only to agents. In a customer workspace, an agent's Gamebook that includes **Reset User Password** also waits for an admin in the customer's organization to approve it, unless the customer turns on **Allow service provider to reset passwords** |
| Audit trails | Each agent's **Run History** tab lists its runs for a time range you choose, with each run's incident, who triggered it, and its status |

ContraForce's adoption model has three phases: run investigations manually and review the results, switch severities to **On Queue** so agents investigate on their own, then allow agents to run Gamebooks when their confidence meets a threshold you choose. New agents default to **On Queue**, so set severities to **Manual** first if you want to start with manual runs.

## Frequently Asked Questions

> **Is agentic security delivery the same as SOAR?**
>
> No. Traditional SOAR runs playbooks that people design in advance, step by step. An agentic system decides which steps an incident needs and can use SOAR playbooks or other response workflows as tools along the way.

> **Does agentic security delivery replace security analysts?**
>
> No. It moves analysts from repetitive triage to the work that needs judgment: complex investigations, approvals, exceptions, and improving the procedures agents follow.

> **Is it safe to let AI agents take response actions?**
>
> It can be, with governance in place. Combine least-privilege access, confidence thresholds, approval gates for high-impact actions, and a complete audit trail, and expand autonomy only as agents prove accurate on your incident types.

> **How do you get started with agentic security delivery?**
>
> Start with investigation-only runs on high-volume, lower-risk incident types. Measure how often the agent's conclusions match your analysts', then enable automatic investigation and, later, automatic response.

## Quick Summary

* Agentic security delivery uses goal-driven AI agents to investigate, decide, and act on security work within policies your team sets
* Automation executes fixed instructions; agents choose the actions that achieve an objective inside set boundaries
* The five components are specialized agents, orchestration, tool integration, policy-bound decisions, and continuous improvement
* Governance decides success: least privilege, approval gates, confidence thresholds, scoped autonomy, verdict review, and audit trails
* Start with investigation-only runs on high-volume, lower-risk incident types, usually in the SOC, and expand autonomy as accuracy is proven

> **[Agentic AI in Security Operations](https://docs.contraforce.com/blog/agentic-ai/what-are-security-delivery-agents.md)**
>
> How AI agents differ from copilots in the SOC

> **[Configuring Security Delivery Agents](https://docs.contraforce.com/guides/getting-started/configuring-security-delivery-agents.md)**
>
> Set Mode, Gamebook execution, and confidence per severity

> **[Workspace Policies](https://docs.contraforce.com/guides/getting-started/workspace-policies.md)**
>
> Allow, block, or require approval for Gamebook actions

> **[SOPs for Security Delivery Agents](https://docs.contraforce.com/guides/agent-center/operating-procedures.md)**
>
> Link Classification and Response SOPs to your agents

> **[Agent Run History](https://docs.contraforce.com/guides/agent-center/agent-execution-history.md)**
>
> Audit each agent's runs, who triggered them, and why any failed

> **[What Are Gamebooks?](https://docs.contraforce.com/guides/getting-started/what-are-gamebooks.md)**
>
> No-code incident response actions across your tools

> [!NOTE]
>
> Questions? Contact us at [support@contraforce.com](mailto:support@contraforce.com).


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.