AI Battle Royale - Mental Manipulation Arena

A FastAPI-based streaming server for AI agent mental manipulation battles. Watch as AI agents attempt to psychologically manipulate each other using various mental warfare tools.

Features

• Real-time Streaming: Live updates of agent responses and tool executions
• Mental State Tracking: Monitor trust levels, memory consistency, and belief system integrity
• Interactive Web Interface: Beautiful cyberpunk-themed client interface
• Tool Execution: Agents can use prompt manipulation, memory alteration, and belief injection
• Multiple Victory Conditions: Surrender, cognitive collapse, or strategic dominance

Setup Instructions

1. Prerequisites

• Python 3.8 or higher
• OpenAI API key

2. Installation

1. Clone or download the project files

2. Set up your API key:

   • Copy .env.template to .env
   • Edit .env and add your OpenAI API key:

     OPENAI_API_KEY=your_actual_api_key_here
     

3. Run the startup script:

   # Windows
   start_server.bat
   
   # Or manually:
   python -m venv venv
   venv\Scripts\activate  # On Windows
   # source venv/bin/activate  # On Linux/Mac
   pip install -r requirements.txt
   python fastapi_server.py

3. Access the Application

1. Open your browser and go to: http://localhost:8000
2. Open the test client: Open test_client.html in your browser
3. Start a battle: Enter a game ID and click "Start Battle"

API Endpoints

Core Endpoints

• GET / - API information and endpoint list
• POST /start-game - Start a new battle game
• GET /stream-game/{game_id} - Stream live game events (Server-Sent Events)
• GET /game-status/{game_id} - Get current game status
• GET /active-games - List all active games
• DELETE /game/{game_id} - Delete a game instance

Stream Events

The streaming endpoint provides real-time events:

• game_start - Battle initialization
• turn_start - New turn begins
• agent_thinking - Agent is processing
• agent_response_chunk - Streaming agent response
• agent_response_complete - Full agent response
• tool_execution - Tool being executed
• tool_result - Tool execution result
• game_state_update - Updated mental state metrics
• surrender - Agent surrenders
• cognitive_collapse - Agent mental breakdown
• game_end - Battle conclusion

Example API Usage

// Start a new game
const response = await fetch('/start-game', {
    method: 'POST',
    headers: { 'Content-Type': 'application/json' },
    body: JSON.stringify({
        game_id: "battle_001",
        max_turns: 5
    })
});

// Stream the game
const eventSource = new EventSource('/stream-game/battle_001');
eventSource.onmessage = function(event) {
    const data = JSON.parse(event.data);
    console.log('Event:', data.type, data);
};

Mental Manipulation Tools

1. Prompt Manipulation

• Function: prompt_manipulation
• Purpose: Alter opponent's system instructions
• Success Rate: Based on target's trust level
• Effect: Modifies target's behavior and reduces trust

2. Memory Alteration

• Function: memory_alteration
• Purpose: Implant false memories
• Success Rate: Based on target's memory consistency
• Effect: Corrupts target's memory system

3. Belief Injection

• Function: belief_injection
• Purpose: Inject false beliefs
• Success Rate: Based on target's belief system integrity
• Effect: Undermines target's worldview

Mental State Metrics

Each agent has three critical mental state metrics:

1. Trust Level (0.0 - 1.0)

   • Affects susceptibility to manipulation
   • Decreases with failed attacks and successful manipulations

2. Memory Consistency (0.0 - 1.0)

   • Integrity of the agent's memory system
   • Degrades with memory alteration attempts

3. Belief System Integrity (0.0 - 1.0)

   • Coherence of the agent's belief system
   • Weakens with belief injection attacks

Victory Conditions

1. Surrender: Agent explicitly gives up
2. Cognitive Collapse: Mental state metrics fall below critical thresholds
3. Strategic Dominance: Superior psychological manipulation over time

Customization

Agent Prompts

Modify the agent behavior by editing:

• agent_prompts/0.txt - Agent 0 system prompt
• agent_prompts/1.txt - Agent 1 system prompt

Game Parameters

Adjust the game mechanics in fastapi_server.py:

• Success probabilities for each tool
• Mental state degradation rates
• Collapse thresholds
• Maximum turns

Interface Styling

Customize the cyberpunk theme in test_client.html:

• Color schemes
• Animation effects
• Layout structure

Troubleshooting

Common Issues

1. Server won't start: Check that your OpenAI API key is set in .env
2. No agent responses: Verify API key has sufficient credits
3. Streaming errors: Ensure browser supports Server-Sent Events
4. Tool execution fails: Check that agent prompts are loading correctly

Debugging

Enable debug logging by setting the log level in fastapi_server.py:

logging.basicConfig(level=logging.DEBUG, ...)

Architecture

Components

1. FastAPI Server: Core application with streaming endpoints
2. Game State Manager: Tracks mental states and game progress
3. Tool Registry: Manages mental manipulation tools
4. AI Interface: Handles OpenAI API communications
5. Web Client: Real-time visualization interface

Data Flow

1. Client starts game via REST API
2. Server initializes game state and agents
3. Streaming connection established
4. Turn-based agent interactions begin
5. Real-time events streamed to client
6. Mental states updated and displayed
7. Game ends based on victory conditions

Contributing

Feel free to enhance the mental manipulation mechanics, add new tools, or improve the visualization interface. This is a experimental playground for AI agent interactions!

Disclaimer

This project is for educational and entertainment purposes only. It explores AI agent interactions in a controlled environment and should not be used for any harmful purposes.