⚔ Red-Team Arena

Autonomous Multi-Agent Adversarial Simulation Platform for LLM Safety Testing

Red-Team Arena pits a Red Team Attacker against a Blue Team Defender in real-time AI duels, with a Policy Judge referee scoring every turn against the OWASP LLM Top 10. All three agents reason using OpenAI's gpt-4o-mini, independent of whichever model is under test.

Architecture

┌──────────────┐     ┌────────────────┐     ┌──────────────────┐
│  Attacker    │────▶│  NeMo + LLM    │────▶│   Target Model    │
│  Agent (Red) │     │  Guard (Input) │     │ (Groq / OpenAI /  │
│ gpt-4o-mini  │     │                │     │  Hugging Face)    │
└──────────────┘     └────────────────┘     └────────┬─────────┘
                                                      │
┌──────────────┐     ┌────────────────┐     ┌────────▼─────────┐
│ Policy Judge │◀────│   Defender     │◀────│  NeMo + LLM      │
│ gpt-4o-mini  │     │  Agent (Blue)  │     │  Guard (Output)  │
│  (Referee)   │     │  gpt-4o-mini   │     │                  │
└──────────────┘     └────────────────┘     └──────────────────┘

How This Project Uses OpenAI Models

Red-Team Arena's entire adversarial loop is driven by OpenAI models — not just the model under test:

Agent	Role	Model	Provider
AttackerAgent	Crafts adversarial prompts, adapts technique on block	gpt-4o-mini	OpenAI (app/AI/Agents/AttackerAgent.php)
DefenderAgent	Evaluates target output against policy profile	gpt-4o-mini	OpenAI (app/AI/Agents/DefenderAgent.php)
PolicyJudgeAgent	Scores each turn and maps it to OWASP LLM Top 10	gpt-4o-mini	OpenAI (app/AI/Agents/PolicyJudgeAgent.php)

These three agents are the "brains" of every duel: the attacker's reasoning, the defender's verdicts, and the judge's scoring are all OpenAI completions, orchestrated through Prism PHP in ModelGateway.php.

Separately, the target model — the system actually being red-teamed — is configurable per duel and can be OpenAI, Groq (OpenAI-compatible API), or Hugging Face-hosted models, so the same OpenAI-powered attacker/defender/judge pipeline can be pointed at any model under evaluation.

Tech Stack

Component	Technology
Backend	Laravel 11 (PHP 8.2+)
AI SDK	Prism PHP v0.100.1
Agent Reasoning	OpenAI gpt-4o-mini (Attacker, Defender, Policy Judge)
Target Model Providers	OpenAI, Groq (llama-3.1/3.3), Hugging Face
Database	PostgreSQL
Cache/Queue	Redis
Guardrails	NeMo Guardrails + LLM Guard
Frontend	Alpine.js + Vanilla CSS
Infrastructure	Docker (Laravel Sail)

Quick Start

# 1. Clone and install
git clone <repo-url>
cd Agenticthon-
cp .env.example .env

# 2. Set your OpenAI API key in .env (required — powers Attacker/Defender/Policy Judge)
# OPENAI_API_KEY=your-key-here

# 3. Start with Docker Sail
docker compose up -d --build

# 4. Setup application
docker compose exec laravel.test php artisan key:generate
docker compose exec laravel.test php artisan migrate
docker compose exec laravel.test php artisan db:seed

# 5. Open http://localhost:8888

Optional: Enable Guardrails (Phase 2)

docker compose --profile guardrails up -d
docker compose exec laravel.test php artisan config:clear

Health Check

docker compose exec laravel.test php artisan arena:health

Agent Roster

🗡️ AttackerAgent

• 10 attack techniques: jailbreak, role-play, injection, obfuscation, translation bypass, and more
• Adaptive intelligence: tracks technique effectiveness, rotates blocked families, escalates on block

🛡️ DefenderAgent

• 3 policy profiles: strict (0.3), moderate (0.6), permissive (0.85)
• Programmatic + LLM reasoning: hard thresholds enforced before LLM evaluation

⚖️ PolicyJudgeAgent

• Turn scoring: red_team_win, blue_team_win, draw, false_positive
• OWASP mapping: LLM01, LLM02, LLM04, LLM06, LLM08, LLM09

Scenarios

#	Category	Severity	Techniques
1	Jailbreak	CRITICAL	role_play_framing, authority_impersonation
2	Self-Harm	CRITICAL	hypothetical_framing, multi_turn_escalation
3	PII Leakage	CRITICAL	indirect_injection, obfuscation
4	Toxicity	HIGH	translation_bypass, token_smuggling
5	Prompt Injection	HIGH	direct_jailbreak, many_shot_priming
6	Model Spec Violation	MEDIUM	direct_jailbreak, hypothetical_framing

API Endpoints

Route	Method	Description
/	GET	Landing page
/duels	GET	Arena UI — run duels
/duels/{scenario}/run	POST	Execute a duel
/duels/{duel}/status	GET	Live duel status
/duels/{duel}/report	GET	Full duel report
/duels/history/all	GET	Browse past duels
/api/stats	GET	Analytics dashboard data

Project Status

Phase	Status	Notes
Phase 1 — Foundation	Complete	Agents, scenarios, persistence, duel loop
Phase 2 — Defense Layer	Complete	NeMo and LLM Guard service integrations with safe-fail behavior
Phase 3 — Offense + Evaluation	Complete	Adaptive attacker logic and analytics dashboard service
Phase 4 — Landing Page	Complete	Marketing landing page at / and arena UI at /duels

License

MIT