Exploit Payload Generator (AI agentic assistant)]

A powerful AI-agentic tool that generates and validates advanced exploit payloads using CrewAI framework.

This project leverages multiple AI agents working together to create, review, and refine exploitation techniques with a focus on EDR evasion and OPSEC considerations.

It follows the methodology of Feedback loop inside agentic workflow.

Problem Statement

Red Team operations often require complex exploit payloads that need to:

• Be technically accurate and effective
• Evade modern EDR solutions
• Maintain OPSEC
• Include proper implementation details
• Be thoroughly validated

Manually developing these payloads is time-consuming and error-prone. This project automates the process using AI agents.

Solution

[Non-Perfect intial PoC

The project implements a multi-agent AI system using CrewAI with:

1. Payload Generator Agent: Creates advanced exploit payloads with:

   • Technical implementation details
   • EDR evasion techniques
   • Build instructions
   • Testing procedures

2. Payload Reviewer Agent: Validates payloads for:

   • Technical accuracy
   • Implementation feasibility
   • OPSEC considerations
   • Documentation completeness

3. Feedback Loop System:

   • Automated refinement through up to 3 iterations
   • Detailed feedback for improvements
   • Quality assurance checks

Sequential Diagram working workflow for this FLOW:

sequenceDiagram
    participant User
    participant ExploitPayloadGeneratorHelperFlow
    participant PayloadExploitCrew
    participant PayloadExploitReviewCrew
    participant FileSystem

    User->>ExploitPayloadGeneratorHelperFlow: kickoff()
    activate ExploitPayloadGeneratorHelperFlow

    rect rgb(200, 220, 240)
        Note over ExploitPayloadGeneratorHelperFlow: Start: generate_PayloadExploit
        ExploitPayloadGeneratorHelperFlow->>PayloadExploitCrew: crew().kickoff(context, feedback)
        PayloadExploitCrew-->>ExploitPayloadGeneratorHelperFlow: PayloadExploit_draft
    end

    rect rgb(220, 200, 240)
        Note over ExploitPayloadGeneratorHelperFlow: Router: evaluate_PayloadExploit
        ExploitPayloadGeneratorHelperFlow->>PayloadExploitReviewCrew: crew().kickoff(PayloadExploit_draft)
        PayloadExploitReviewCrew-->>ExploitPayloadGeneratorHelperFlow: {valid, feedback}
        
        alt is valid
            ExploitPayloadGeneratorHelperFlow->>ExploitPayloadGeneratorHelperFlow: route to "completed"
        else retry count < 3
            ExploitPayloadGeneratorHelperFlow->>ExploitPayloadGeneratorHelperFlow: route to "retry"
            ExploitPayloadGeneratorHelperFlow->>PayloadExploitCrew: Retry with feedback
        else retry count >= 3
            ExploitPayloadGeneratorHelperFlow->>ExploitPayloadGeneratorHelperFlow: route to "max_retry_reached"
        end
    end

    alt completed
        ExploitPayloadGeneratorHelperFlow->>FileSystem: Save PayloadExploit.md
        FileSystem-->>ExploitPayloadGeneratorHelperFlow: File saved
    else max_retry_reached
        ExploitPayloadGeneratorHelperFlow->>User: Display final PayloadExploit and feedback
    end

    deactivate ExploitPayloadGeneratorHelperFlow

Requirements & Local Setup

• Python >=3.10 < 3.13
• UV package manager
• Ollama (for local LLM support)

Local Setup

1. Clone the repository:

git clone https://github.com/Abhinandan-Khurana/exploit-payload-generator-ai-agent.git
cd exploit-payload-generator-ai-agent

2. Install UV if not already installed:

pip install uv

3. Install dependencies:

uv pip install -e .
# OR
crewai install

4. Install Ollama for local LLM support:

curl https://ollama.ai/install.sh | sh

5. Install whiterabbitneo and hermes3:

ollama pull rfc/whiterabbitneo
ollama pull hermes3

Running the Project

1. Start Ollama (if using local LLMs):

ollama serve

2. Run the project

crewai flow kickoff

The system will generate a payload, review it, and either:

• Save the final payload to generated_final_payload.md if validated
• Display feedback and retry up to 3 times if issues are found

Important Notes

• Results may be redundant or vary between runs due to the nature of LLM responses
• The system prioritizes accuracy over speed
• Each iteration improves payload quality through AI feedback
• Local LLMs may provide different results compared to OpenAI's models

Contributing

Contributions are welcome! Please feel free to submit a Pull Request. For major changes, please open an issue first.

Author

Abhinandan Khurana