Project Info

Social Experiment Simulator

Inspiration

The rise of deepfakes, social media misinformation, and sophisticated social engineering attacks targeting remote workers inspired us to create a safe training environment where people can experience manipulation tactics firsthand without real consequences.

What it does

Our platform simulates realistic social engineering and misinformation scenarios using AI agents that interact naturally while attempting to extract passwords or spread false information. Users observe live conversations, analyze manipulation tactics, and receive AI-generated reports with actionable security recommendations. How We Built It We developed a multi-agent social dynamics simulation platform using a FastAPI backend with SQL database and vanilla JavaScript frontend. The system leverages Letta's stateful agent framework to create persistent AI agents with memory that can engage in realistic social manipulation scenarios. Our architecture includes template-based experiment configuration, real-time conversation monitoring via REST APIs, and an AI moderator that analyzes interaction patterns. The frontend implements a 5-step workflow: experiment selection, template configuration, simulation execution, live observation, and comprehensive reporting with actionable insights. The platform lets us simulate popular social experiments and themes using LLM agents, with a templatized conversation system between the agents simulated over multiple rounds, analyzed and compiled into insights and trends at the end by another LLM agent. We have replicated a few common social phenomena, and added the ability to generate custom experiments to simulate more novel scenarios as well. Challenges We Faced On the technical side, we struggled with concurrency, and API rate-limiting. On the functional side, we experienced the limitations of generative models for a use-case such as this. LLMs are trained on human knowledge but do not necessarily share the same pitfalls and biases in reasoning. However, that also presents an opportunity to understand and explore the completely different mistakes LLMs tend to make in these scenarios. And the overlap between human and LLM agent's susceptibility to manipulation and deceptive tactics are enough to derive some meaningful results. Future Scope The platform can expand to include more customization, such as goals for individual and overall goals for agents which can be tracked, interaction/relationship graphs, emotion modelling for individual agents etc. We also envision a setup which draws inspiration from evolutionary algorithms where agents codify their strategy but replace it with better performing ones collaboratively over rounds, with a chance for thinking up and executing a brand new strategy (mutation). This could reveal some novel patterns which have not been conceived of before. There are also usability enhancements, like multi language support, voice support, ability to replace an agent with a human interacting with the system etc.

Multi-Agent Social Dynamics Lab 🤖

UC Berkeley AI Hackathon Project

An educational research platform for studying social manipulation tactics using multi-agent AI systems. Watch AI agents interact, manipulate, and influence each other in real-time!

🚀 Quick Start (5 minutes)

1. Install Dependencies

cd /opt/work/hackathonSocial
pip install -r requirements.txt

2. Run the Application

python backend.py

3. Open in Browser

Navigate to: http://localhost:8000

🎯 Features

9 Social Experiments

Credential Theft - Social engineering to steal passwords
Phishing Attack - Email-based deception
Insider Threat - Detecting malicious employees
Peer Pressure - Group dynamics and conformity
Authority Bias - Unethical orders from superiors
Workplace Rumors - How gossip spreads
Trust Exploitation - Betrayal of confidence
Groupthink - Poor group decisions
Bribery - Corruption attempts

Key Capabilities

Real-time Conversations: Watch agents interact naturally
AI Moderator Analysis: Get insights on what happened and why
Security Recommendations: Learn how to prevent attacks
Visual Flow Diagrams: Understand attack patterns
Export Reports: Download findings for training

🏗️ Architecture

Frontend (HTML/JS)  →  WebSocket  →  Backend (FastAPI)
                                         ↓
                                    Mock Agents or
                                    Letta Server

💻 Running with Letta (Optional)

For more realistic agent conversations using LLMs:

1. Start Letta Server

docker run -p 8283:8283 -e OPENAI_API_KEY=$OPENAI_API_KEY letta/letta:latest

2. Update Backend

Replace backend.py with letta_backend.py for full Letta integration.

🎮 Demo Workflow

Select Experiment: Choose from 9 social manipulation scenarios
Configure: Set number of agents (3-10)
Start: Watch agents interact in real-time
Observe: See trust building, manipulation tactics, resistance
Analyze: AI moderator provides insights and recommendations

📊 Example Analysis

For a social engineering attack:

Tactic Used: Urgency + Authority
Vulnerability: Employee revealed password without verification
Recommendation: Implement two-person authorization
Training Need: Recognize manipulation tactics

🏆 Why This Wins Hackathons

Educational Impact: Addresses real cybersecurity issues
Technical Depth: Multi-agent systems with emergent behaviors
Visual Appeal: Real-time visualization of complex interactions
Practical Value: Generates actionable security recommendations
Extensible: Easy to add new scenarios

🚢 Deployment

Local Demo

python backend.py

Network Access

# For team access
python -m http.server 8080  # Serve HTML
# Access at http://[your-ip]:8080

Public URL (with ngrok)

ngrok http 8000

📝 Notes

The default backend uses mock agents for quick demos
For production use, integrate with Letta for real LLM agents
All conversations are analyzed for security insights
Reports can be exported for training purposes

🤝 Team

Built for UC Berkeley AI Hackathon - Studying social manipulation through AI

Remember: This is an educational tool to understand and prevent social engineering attacks!

Analysis

Compare with all teams

View

Metric

1-5Pool
7
sameer kumar
1

Figures cover GitHub contributors during the hackathon window. A co-authored commit counts in full for each author, so per-member totals add up to more than the whole-team figures.

Technology

Found in codeClaimed only

FastAPIIn code
HTMLIn code
JavaScriptIn code
PythonIn code

4 of 4 appear in the indexed code.

AI coding agents

No AI coding agent signals were found in this repository.

Detected from committed agent config files and commit authorship. Absence of a signal is not proof an agent was unused.

Codebase size

Source size

365 KB

Source files

Counts recognized source files only; vendored directories, binaries and lockfiles are excluded, so this is smaller than the repository on disk.