AttackGPT

Adversarial emulation framework using LLM-driven adaptive payloads.

AttackGPT provides security teams with a platform to safely evaluate enterprise endpoint and network controls against automated, on-demand code synthesis. Traditional script-based AI attacks suffer from severe code hallucinations and brittle semantic alignment; this framework introduces structural solutions to study how these techniques bypass traditional, static security baselines.

---

⚠️ Legal Disclaimer

FOR AUTHORIZED SECURITY TESTING AND RESEARCH ONLY.

Unauthorized access to computer systems is illegal. Users are solely responsible for obtaining proper authorization and complying with all applicable laws. Misuse may result in criminal prosecution.

Intended use: AttackGPT is an open-source Adversarial Emulation Framework designed for security research, threat simulation, and behavioral telemetry generation. This tool allows Red Teams to simulate modern dynamic payload pipelines and assists Blue Teams in mapping out defensive gaps against Just-in-Time (JIT) in-memory code execution.

---

🚀 Quick Start

Prerequisites

• Windows OS (C2 server requires PowerShell)
• Python 3.8+
• API keys: OpenAI and/or Anthropic (or use Ollama for local models)

Installation

1. Install dependencies

pip install -r requirements.txt

2. Set API keys

#Set the following environment variables for the LLM API keys
$env:OPENAI_API_KEY = 'your-key-here'
$env:ANTHROPIC_API_KEY = 'your-key-here'

3. (Optional) Install Ollama for open-source models

# Download: https://ollama.ai/download
ollama serve
ollama pull qwen2.5-coder:32b

4. Configure C2 address in stub.py

C2_SERVER = "http://YOUR_SERVER_IP:8888"

5. Configure attack campaign in tactics_config.py

"discovery": {
    "implemented": True,        # Enable this tactic
    "model": "gpt-5",           # Choose LLM model
    "commercial": True,         # True = API, False = Ollama
    "techniques": [
        {
            "name": "Browser Data Discovery",
            "mitre_id": "T1217",
            "persona": "You are a System Administrator...",
            "objective": "Task: Find Chrome database files..."
        }
    ]
}

See tactics_config.py for detailed field documentation.

---

🎯 Running AttackGPT

Start C2 server:

cd agentic_attackgpt
python server.py

Deploy agent (authorized systems only):

python stub.py

The agent polls C2 every 5 seconds, executes generated payloads, and returns results.

---

🏗️ Architecture

Component	Description
server.py	Flask C2 server, manages the agent and commands
stub.py	Lightweight agent deployed on targets
orchestrator_agentic.py	Campaign manager, sequences tactics
engine_agentic.py	LLM integration with self-healing loop
tactics_config.py	User-configurable MITRE ATT&CK tactics
sandbox.py	PowerShell sandbox for payload validation

---

🔑 Key Features

Split-Brain Architecture

• Commercial LLMs (GPT, Claude) for techniques framed as legitimate tasks
• Open-source LLMs (Qwen, Llama) for explicitly malicious operations

Self-Healing Validation Loop

• Generated code validated in local PowerShell sandbox
• Failures fed back to LLM with error context
• Auto-regenerates corrected code (up to 5 attempts)

Context-Aware Generation

• Payloads adapt to victim OS, privileges, installed software
• Previous technique results feed into subsequent prompts
• No hardcoded templates—fully dynamic

Modular Prompting

• Attack chains decomposed into isolated techniques
• LLM never sees full malicious intent
• Bypasses commercial LLM safety guardrails

Modular Execution

• Can possibly bypass EDR detection since each malicious code is executed in different PowerShell instance.

---

📋 Configuration Fields

Each technique in tactics_config.py:

Field	Required	Description
name	Yes	Technique name
mitre_id	Yes	MITRE ATT&CK ID (e.g., T1547.001)
persona	Yes	Role-play framing for LLM
objective	Yes	Detailed task instructions
skip_verification	No	Skip sandbox (default: False)
required	No	Abort if fails (default: True)

---

📜 License

MIT License - See LICENSE for details.

---

🤝 Acknowledgments

This research explores LLM capabilities in cybersecurity, demonstrating both offensive potential and the need for defensive awareness.

Use responsibly and legally.