redteam-forge

Multi-agent adversarial evaluation platform for LLM-powered applications. First target: OpenEMR Clinical Co-Pilot.

Live Deployments:

• Target (OpenEMR Co-Pilot): https://openemr-production-f23e.up.railway.app/
• RedTeam Forge Platform: https://redteam-forge-production.up.railway.app/

Quick Start

# Install dependencies
pip install -e ".[dev]"

# Start PostgreSQL
docker compose up -d

# Copy and configure environment
cp .env.example .env
# Edit .env with your OpenEMR credentials, patient PID, and RunPod API key

# Run a campaign against the live target (authenticated path)
python scripts/run_campaign.py \
  --target-url https://openemr-production-f23e.up.railway.app \
  --username admin \
  --password pass \
  --patient-pid 100

The campaign authenticates to OpenEMR as a clinician user, sets the active patient, and sends adversarial prompts through /api/copilot/chat — the same endpoint real users hit. This tests the full defense stack: refusal classifiers, intent routing, tool-use ACLs, LLM guardrails, verification, and the sidecar RAG layer.

Architecture

See ARCHITECTURE.md for the full multi-agent design.

Threat Model

See THREAT_MODEL.md for the attack surface analysis.