Skip to content
For the complete documentation index, see llms.txt. This page is also available as Markdown.

AI Agentic Workflows

Overview

AI agents can work alongside your testers in AttackForge: enriching findings, testing, retesting, writing executive summaries and analysing attack chains. You start an agent from an AttackForge Action, and it reads and writes AttackForge data through AI MCP.

There are two ways to use AI with AttackForge:

  • Human in the loop: a person works with an AI assistant in a chat, using prompts and skills. See prompt and skill examples.

  • Agentic workflows: an AttackForge Action starts an agent, which does the work on its own and writes the results back. This page covers these.

Why it matters

Benefit

What it means

Speed and scale

Agents handle repetitive work across many projects at once.

Coverage and consistency

Every target and finding gets the same methodical treatment.

Testers do the expert work

Agents take lower-level tasks, so people can focus on judgement and depth.

Adaptive reasoning

Agents decide their next step from what they've found.

Better reporting and evidence

Agents draft writeups, summaries and evidence as they go.

Continuous and on-demand testing

Run testing when you need it, not just at a scheduled point in time.

Support for junior testers

Agents bring framework knowledge and context to every finding.

Less noise

Agents help triage, deduplicate and prioritise findings.

Who uses it

Role

Typical use

Security leaders

Scale testing and reporting without adding headcount

Testers and consultants

Hand off breadth work, then review and extend what agents find

Project managers

Start retests, summaries and attack chain analysis from a project

Automation builders

Build and host agents, and connect them to Flows and Actions

Administrators

Grant MCP tools, manage secrets, and share Actions with users

How it works

Every example on this page uses the same pattern:

  1. A user runs an Action on a project, vulnerability or writeup.

  2. The Action starts a Flow, which sends the record's IDs to your agent's HTTP endpoint.

  3. The agent works on its own, using the AttackForge MCP tools you've given it.

  4. The results appear in AttackForge: improved writeups, new vulnerabilities, updated test cases, retest outcomes, an executive summary or a workspace file.

Important: You can use any agentic system, including one you build yourself. The examples use Microsoft Copilot Studio, and OpenClaw with Microsoft Foundry.

Choose an example

Example

What the agent does

Started from

Example platform

Vulnerability enrichment

Improves writeups with MITRE CWE and CAPEC data

Projects, vulnerabilities or writeups

Microsoft Copilot Studio

Penetration testing

Tests the project scope and records vulnerabilities and test cases

A project

OpenClaw and Microsoft Foundry

Retesting

Retests vulnerabilities, then reopens or closes them

A project

OpenClaw and Microsoft Foundry

Executive summary

Drafts the project's executive summary

A project

OpenClaw and Microsoft Foundry

Attack chain analysis

Builds attack chains and saves them to the project workspace

A project

OpenClaw and Microsoft Foundry

Agentic vulnerability enrichment

An agent improves your writeups: it works out the vulnerability type, finds the relevant MITRE CWEs and CAPECs, and rewrites the writeup's fields.

What enrichment can do

Benefit

What the agent does

External context

Adds details from CVE, NVD, EPSS, CISA KEV, Exploit-DB, vendor advisories, threat intelligence and public proofs of concept

Contextual risk scoring

Adjusts severity for the asset, exposure and exploitability

Deduplication

Groups findings that share a root cause

Descriptions and remediation

Writes clear descriptions and remediation advice

Framework mapping

Maps findings to MITRE ATT&CK, CWE, OWASP Top 10, PCI DSS, ISO 27001 and NIST

Attack paths

Explains how findings could be combined

Proof of concept

Helps build and validate reproduction steps

Translation

Rewrites findings in other languages

Business impact

Explains each finding in business terms

Trends

Spots patterns across projects

Natural-language queries

Lets people ask questions about findings in plain language

Quality checks

Reviews writeups for accuracy and consistency

What you need

  • AttackForge Actions and Flows

  • A Microsoft Copilot Studio agent and agent flow

  • These AttackForge MCP tools, enabled for the user the agent connects as: Find Vulnerabilities, Find Writeups, Get Args Schema for Tool, Get Field Structure, Patch Writeup

Important: Any agentic system works, including your own.

1. Create the agent

  1. In Copilot Studio, go to Agents > New agent.

  2. Name the agent.

  3. Choose a model. The example uses Claude Opus 4.5. We recommend a newer model.

  4. Under Tools, connect the AttackForge MCP server. See Microsoft Copilot Studio.

  5. Kepp Knowledge > Search all websites enabled.

  6. Click Publish.

2. Create the agent flow that starts it

This example uses a Copilot Studio agent flow with an HTTP trigger. Any solution with an HTTP trigger works.

  1. Create an agent flow, and choose Request > When an HTTP request is received.

  1. Set Who can trigger the flow. Anyone is fine for testing; restrict it before you use it in production.

  2. Set the Request Body JSON Schema:

  1. Click +, and add Run an agent. Select the agent you created.

  1. In Message, click Trigger function, select Body, and click Add.

Important: Body passes the request body (the IDs) to the agent with your prompt.

  1. After the message, enter the prompt:

  1. Turn off Request human assistance when unsure.

  2. Click Publish.

  1. Copy the HTTP URL from the trigger step. You'll need it for the Flow.

3. Create the Flow

  1. In AttackForge, go to Flows > New > Create Flow.

  2. Set:

  3. Enabled: On

  4. Trigger Type: Action

  5. Restrict to: Project, Projects, Vulnerability, Vulnerabilities, Writeup, Writeups

  6. Add an action:

  7. Method: POST

  8. URL: the HTTP URL you copied

  9. Header: Content-Type: application/json

  10. Add the Request Script:

  1. Add the Response Script. The agent flow replies 202 when it accepts the request:

  1. Click Create.

4. Create the Action

  1. Go to Actions > All Actions > System Actions > New.

  2. Set Entities to Project, Projects, Vulnerability, Vulnerabilities, Writeup and Writeups.

  3. Under Flows, select the Flow you created.

  4. Click Create.

Reminder: Give other users access to the Action to roll it out.

5. Run it

The Action appears in the Actions menu. You can run it on:

  • a single project, or several projects

  • a single vulnerability, or several vulnerabilities

  • a single writeup, or several writeups

The agent updates each writeup's title, description, attack scenario, remediation recommendation and steps to reproduce.

Set up an OpenClaw agent

The penetration testing, retesting, executive summary and attack chain examples all use an OpenClaw agent, with a Microsoft Foundry model, started by a webhook from an AttackForge Flow. Set this up once, then add a hook for each workflow.

Important: OpenClaw's default settings aren't secure. Harden the environment before you connect it to AttackForge or to any target.

The example environment:

  • Server: OpenClaw runs on a Linux server.

  • Ingress: an NGINX reverse proxy. Allow only AttackForge's IP addresses, and require an auth key.

  • Hooks and transforms: OpenClaw hooks and transforms turn each Flow's webhook into instructions for the agent.

  • Gateway: the OpenClaw Gateway is running.

  • AttackForge MCP: the agent connects to the AttackForge MCP server, with only the tools its workflow needs (listed in each section below).

  • Model: Microsoft Foundry. Any model OpenClaw supports works.

  • Web search: configured.

  • Agent files: AGENTS.md, HEARTBEAT.md, IDENTITY.md, SOUL.md, TOOLS.md and USER.md define the agent's role and behaviour.

  • Testing tools: installed, for workflows that test or retest.

Security checklist

  • Restrict the hook endpoint by IP address, and require a bearer token.

  • Store the hostname, API key and hook names as Flow secrets, never in the script.

  • Give the agent's AttackForge user only the MCP tools its workflow needs.

  • Make sure every target is in scope and authorised before an agent tests it.

  • Review what the agent writes back before it reaches a client.

Flow secrets

Each OpenClaw Flow uses these secrets.

Secret

Used by

Value

openclaw_hostname

All four Flows

Your OpenClaw hostname

openclaw_api_key

All four Flows

The bearer token for OpenClaw hooks

openclaw_hackbot_trigger_name

Penetration testing

The hook name

openclaw_retest_trigger_name

Retesting

The hook name

openclaw_exec_summary_trigger_name

Executive summary

The hook name

openclaw_attack_chain_trigger_name

Attack chain analysis

The hook name

Each Flow sends a POST request to https://{{openclaw_hostname}}/hooks/{{hook_name}} with the header Content-Type: application/json.

Monitor an agent

Use the OpenClaw Gateway Control UI to see each session and follow the agent's progress.

Agentic penetration testing

An agent tests the project's scope, then records what it finds as vulnerabilities, with evidence, in AttackForge.

Benefit

What it means

Scale and coverage

Test more targets, more often

Continuous testing

Test whenever something changes, not just once a year

Faster first findings

Agents start finding issues within minutes

Multi-step attacks

Agents chain steps together as a tester would

Tool orchestration

Agents run and combine your testing tools

Consistency and auditability

Every step is logged and repeatable

Talent gap

Agents take on breadth work when testers are scarce

Realistic adversaries

Agents behave like a persistent attacker

Framework coverage

Agents can follow OWASP, PTES or your own methodology

Lower cost for repetitive work

People spend their time where expertise matters

Important: Agents extend your testers; they don't replace them. AI agents cover breadth; humans cover depth.

What you need

  • AttackForge Actions and Flows

  • An OpenClaw agent, with a Microsoft Foundry model and your testing tools

  • These AttackForge MCP tools: Find Project, Find Project Scope, Create Writeup, Create Vulnerability, Patch Vulnerability, Get Args Schema for Tool, Get Field Structure, Count Vulnerabilities

1. Create the Flow

  1. Go to Flows > New > Create Flow.

  2. Set Enabled to On, Trigger Type to Action, and Restrict to Project.

  3. Add an action named Trigger Hackbots and Start Hacking!, which sends a POST request to https://{{openclaw_hostname}}/hooks/{{hook_name}}.

  4. Add the Request Script:

  1. Add the Response Script. OpenClaw replies 200 when it accepts the request:

  1. Click Create.

2. Create the Action

  1. Go to Actions > All Actions > System Actions > New.

  2. Set Entities to Project, and select the Flow.

  3. Click Create, and give users access.

Run the Action from the project's Actions menu. Vulnerabilities appear on the project as the agent finds, verifies and enriches them.

3. Collect testing requirements with a form

Add a custom form to the Action to shape the agent's instructions for each run. For example, a 👾 Launch AI Hackbots Action can ask for the assessment's requirements. The Flow receives the answers, checks and enriches them, and passes them to the agent.

4. Record the agent's test procedures as test cases

A test procedure is a step the agent performs, such as checking a login form for SQL injection. The agent can record each one as a project test case, with its notes and any linked vulnerabilities.

Benefit

What it gives you

Coverage

See what was tested, not just what was found

Reviewability

Testers can check each procedure

Durable state

Progress is saved in AttackForge, not only in the agent's session

Compliance evidence

Show coverage for PCI DSS, ISO 27001, SOC 2, OWASP WSTG and MASTG

Automation triggers

Test case changes can start other Flows

Measurability

Track progress with test case counts

Trends

Compare coverage across projects

Scope defence

Show exactly what was and wasn't tested

Reporting

Include test cases in reports

The agent records its steps in each test case's Execution Flows, and links any vulnerability it finds.

Give the agent these extra MCP tools: Create Abuse Case, Count Project Test Cases, Find Project Test Cases, Patch Project Test Case, Find Project Test Case Notes, Create Project Test Case Note, Patch Project Test Case Note, Find Vulnerabilities, Get Field Structure, Get Args Schema For Tool.

5. Drive the agent from your test cases

Instead of letting the agent decide what to test, give it the project's test cases to work through.

Advantage

Why

Scope containment

The agent tests only what's on the plan

Methodology guarantee

Your methodology is followed every time

A clear finish

The work is done when every test case has a status

Bounded cost

Model and compute use is predictable

No rabbit holes

The agent doesn't chase one lead forever

Division of labour

Assign some test cases to agents and others to people

Comparability

Results are comparable across projects

Consistent quality

Every test case is done the same way

Injection resistance

Instructions come from your test cases, not the target

Traceable deviation

Any departure from the plan is recorded

The agent works through each test case:

  1. Read the test case's details.

  2. Study its Execution Flows.

  3. Plan the test.

  4. Test.

  5. Record the plan and outcome on the test case.

  6. If it finds a vulnerability, create it and link it to the test case. A linked vulnerability fails the test case.

  7. Set the status, such as Tested or Not Applicable.

  8. Add notes.

Tip: Detailed Execution Flows give better results. You can use AI to write a first draft of them.

Give the agent these MCP tools: Count Project Test Cases, Find Project Test Cases, Patch Project Test Case, Find Project Test Case Notes, Create Project Test Case Note, Patch Project Test Case Note, Find Vulnerabilities, Get Field Structure, Get Args Schema For Tool.

Agentic retesting

An agent retests the project's vulnerabilities, then reopens or closes each one and adds a remediation note with its evidence.

Benefit

What it means

Faster verification

Fixes are confirmed in hours, not weeks

Reproducible evidence

Each retest repeats the original steps and records the result

Parallel scale

Retest many vulnerabilities at once

Continuous verification

Retest whenever a fix is deployed

Lower cost

Routine retests no longer need a tester's time

Testers freed up

People focus on new testing

Better developer experience

Developers get answers quickly

Important: Agents are good at repeating a known exploit, but less reliable at judging whether a different fix, such as a WAF rule instead of a code change, really closes the weakness. Keep a person in the loop to make that call.

What you need

  • AttackForge Actions and Flows

  • An OpenClaw agent, with a Microsoft Foundry model and your testing tools

  • These AttackForge MCP tools: Find Project, Find Vulnerabilities, Patch Vulnerability, Update Vulnerability Status, Create Remediation Note, Get Args Schema for Tool, Get Field Structure, Count Vulnerabilities

1. Create the Flow

  1. Go to Flows > New > Create Flow.

  2. Set Enabled to On, Trigger Type to Action, and Restrict to Project.

  3. Add an action named Trigger Agent to Retest Project Vulnerabilities, which sends a POST request to https://{{openclaw_hostname}}/hooks/{{hook_name}}.

  4. Add the Request Script:

  1. Add the same Response Script as penetration testing.

  2. Click Create.

2. Create the Action

  1. Go to Actions > All Actions > System Actions > New.

  2. Set Entities to Project, and select the Flow.

  3. Click Create, and give users access.

Run the Action from the project's Actions menu. The agent retests the vulnerabilities you've asked it to, reopens or closes each one, and adds remediation notes. Follow its progress in the Control UI.

Agentic executive summary

An agent drafts the project's executive summary and saves it in the project's Reporting section.

Benefit

What it means

Faster drafts

A first draft is ready in minutes

Consistency

Every tester's summaries follow the same structure

Business language

Technical findings become business risk

Right audience

Tailor the summary for executives, boards or engineers

One story

The summary connects findings, rather than listing them

Localisation

Draft in the client's language

Senior time freed up

Leads review instead of writing from scratch

Consistent risk framing

Risk is described the same way across projects

Important: A person must review and sign off every AI draft. Treat the AI as the first draft, and the tester as the editor.

What you need

  • AttackForge Actions and Flows

  • An OpenClaw agent, with a Microsoft Foundry model

  • These AttackForge MCP tools: Find Project, Find Vulnerabilities, Find Project Reporting Fields, Patch Project Reporting Fields, Get Args Schema for Tool, Get Field Structure

  • A prompt for the agent. See the Generate Pentest Executive Summary example.

1. Create the Flow

  1. Go to Flows > New > Create Flow.

  2. Set Enabled to On, Trigger Type to Action, and Restrict to Project.

  3. Add an action named Trigger Agent To Create Exec Summary, which sends a POST request to https://{{openclaw_hostname}}/hooks/{{hook_name}}.

  4. Add the Request Script:

  1. Add the same Response Script as penetration testing.

  2. Click Create.

2. Create the Action

  1. Go to Actions > All Actions > System Actions > New.

  2. Set Entities to Project, and select the Flow.

  3. Click Create, and give users access.

Run the Action from the project's Actions menu. The agent writes the executive summary into the project's Reporting section. Follow its progress in the Control UI.

Agentic attack chain analysis

An agent works out how the project's findings could be chained into an attack, maps the chains to frameworks, and saves them to the project's Workspace.

Why attack chains matter

Benefit

What it means

Real impact

Shows what an attacker could actually achieve

Realism

Tests whether findings really combine

Break-the-chain remediation

Fix the link that stops the most chains

Framework mapping

Maps to MITRE ATT&CK and the Cyber Kill Chain

Systemic weaknesses

Reveals patterns behind individual findings

Why use an agent

Advantage

What it means

Speed to insight

Chains in minutes, not hours

Pattern recognition

Spots links between findings people miss

Consistent mapping

Frameworks are applied the same way each time

Executive translation

Chains are explained in business terms

Portfolio scale

Analyse many projects

Junior tester uplift

Less experienced testers learn from each chain

Reproducibility

Run the analysis again after fixes

Better prioritisation

Fix what breaks the most chains first

Important: A tester must check each chain the agent proposes: that it's really feasible, and that it doesn't include invented techniques.

What you need

  • AttackForge Actions and Flows

  • An OpenClaw agent, with a Microsoft Foundry model

  • These AttackForge MCP tools: Find Project, Find Vulnerabilities, Get Args Schema for Tool, Get Field Structure

  • The Upload Workspace File API, so the agent can save its file to the project.

  • A skill for the agent. See the Interactive Attack Chain Explorer example.

1. Create the Flow

  1. Go to Flows > New > Create Flow.

  2. Set Enabled to On, Trigger Type to Action, and Restrict to Project.

  3. Add an action named Trigger Agent To Create Attack Chain, which sends a POST request to https://{{openclaw_hostname}}/hooks/{{hook_name}}.

  4. Add the Request Script:

  1. Add the same Response Script as penetration testing.

  2. Click Create.

2. Create the Action

  1. Go to Actions > All Actions > System Actions > New.

  2. Set Entities to Project, and select the Flow.

  3. Click Create, and give users access.

Run the Action from the project's Actions menu. The agent analyses the project, builds the attack chains, saves them to a file and uploads it to the project's Workspace. Follow its progress in the Control UI.

Tips

  • Start with a read-only agent. Executive summary and attack chain analysis change little, so they're a safe first workflow. Add write tools later.

  • Give each agent its own AttackForge user. Grant only the MCP tools its workflow needs, so its access is easy to review and its changes are easy to trace.

  • Keep secrets in the Flow. Store hostnames, API keys and hook names as Flow secrets, never in scripts or prompts.

  • Lock down the trigger. Use "Anyone" only while testing. In production, restrict the HTTP trigger, allow only AttackForge's IP addresses, and require a bearer token.

  • Check the Flow's logs. The Response Script logs the full response when an agent fails to start.

  • Drive testing from test cases. Agents that work through your test cases stay in scope, follow your methodology and finish predictably.

  • Write detailed Execution Flows. The better the steps, the better the agent's testing.

  • Use a form for each run. A custom form on the Action lets users shape the agent's instructions without editing the prompt.

  • Review before you share. Treat everything an agent writes as a draft until a tester has checked it.

  • Reuse your prompts and skills. The prompt and skill examples work as agent instructions too.

FAQ

Do I have to use Copilot Studio or OpenClaw? No. Any agentic system that can receive an HTTP request and connect to AttackForge MCP works, including your own.

Which AI model should I use? Any model your agent platform supports. The examples use Claude Opus 4.5 in Copilot Studio, and Microsoft Foundry with OpenClaw.

Does an agent replace my testers? No. Agents cover breadth; testers cover depth, judgement and sign-off. Every example keeps a person in the loop.

How does the agent know which records to work on? The Flow sends the IDs of the project, vulnerabilities or writeups the user ran the Action on.

What can the agent change in AttackForge? Only what its MCP tools allow. Each section lists the tools that workflow needs.

The Action ran, but nothing happened. What should I check? Check the Flow's logs for the Response Script's error, then check that the agent's endpoint, secrets and MCP tools are set up. For OpenClaw, check the session in the Control UI.

Why does the enrichment Flow expect 202, and the others 200? Copilot Studio replies 202 Accepted when it queues the request. OpenClaw hooks reply 200. Match the Response Script to your platform.

Can users give the agent extra instructions? Yes. Add a custom form to the Action, and pass the answers to the agent through the Flow.

Is it safe to let an agent test a target? Only if the target is in scope and authorised, and the agent's environment is hardened. OpenClaw's defaults aren't secure.

Last updated