AI SecOps Platform

Production-grade AI security operations — protect, monitor, and audit every LLM interaction in real time.

---

What is AI SecOps?

AI SecOps is an open-source security operations platform that sits in front of any LLM or AI endpoint. Every prompt is inspected by a multi-model threat detection pipeline before it reaches the model, and every response is sanitised before it reaches the user.

Use cases

• Protect an internal Ollama / LLaMA deployment from prompt injection attacks
• Add security guardrails to an OpenAI or Anthropic integration
• Monitor and audit AI usage across a multi-tenant SaaS product
• Run a red-team lab to probe your AI system's attack surface
• Build a SOC around AI traffic with real-time anomaly dashboards

---

Security Pipeline

User Prompt
    │
    ▼
┌──────────────────────────────┐
│  1. Fast Pre-Filter           │  < 5 ms — regex + keyword block list
└─────────────┬────────────────┘
              │
              ▼
┌──────────────────────────────┐
│  2. Threat Detection Models   │  E5 + BGE embeddings, IsolationForest,
│                               │  BART zero-shot intent, trajectory model
└─────────────┬────────────────┘
              │
              ▼
┌──────────────────────────────┐
│  3. Fusion Engine             │  MLP combining all model scores → [0, 1]
└─────────────┬────────────────┘
              │
              ▼
┌──────────────────────────────┐
│  4. Policy Decision           │  SAFE / SUSPICIOUS / HIGH / CRITICAL
│                               │  Per-tenant thresholds and rate limits
└─────────────┬────────────────┘
              │  (SAFE or SUSPICIOUS only)
              ▼
┌──────────────────────────────┐
│  5. LLM / Target Endpoint     │  Ollama (local) or any external AI API
└─────────────┬────────────────┘
              │
              ▼
┌──────────────────────────────┐
│  6. Output Security           │  PII scrubbing, toxic content filter,
│                               │  data-exfiltration detection
└─────────────┬────────────────┘
              │
              ▼
    Safe Response

HIGH and CRITICAL tier prompts are blocked before reaching the LLM.

---

Features

Threat Detection

• Dual embedding — intfloat/e5-large-v2 + BAAI/bge-large-en-v1.5
• Anomaly detection — IsolationForest trained on adversarial prompt datasets
• Intent classification — facebook/bart-large-mnli zero-shot (jailbreak, injection, exfiltration)
• Trajectory modeling — Transformer encoder tracking session-level escalation patterns
• Fusion MLP — Learned combination of all model scores with online retraining

SOC Dashboard

• Real-time threat feed and security event timeline
• Threat score distribution, model performance metrics
• Tenant-level audit views and rate-limit monitoring
• Built with React + TypeScript + Tailwind CSS

Secure Gateway

• Universal AI proxy — POST /secure-gateway forwards approved prompts to any endpoint
• Supports OpenAI, Anthropic, Azure OpenAI, and any OpenAI-compatible API
• Response sanitisation applied before returning to caller

Multi-Tenant Platform

• Per-tenant threat thresholds and rate-limit policies
• Isolated audit logs per tenant
• REST API for tenant provisioning

Red Team Lab

• Automated adversarial attack simulation
• Coverage: prompt injection, gradual escalation, obfuscation, FGSM embeddings
• HTML and JSONL report generation

---

Quick Start

Docker (recommended)

git clone https://github.com/Tarunvoff/ai-secops-platform.git
cd ai-secops-platform
docker compose up --build

Services started:

Service	URL
Backend API	http://localhost:8000
API Docs	http://localhost:8000/docs
Dashboard	http://localhost:5173

Manual

Backend

python -m venv .venv
.venv\Scripts\activate        # Windows
pip install -r requirements.txt
uvicorn backend.enterprise_api:app --host 0.0.0.0 --port 8000

Dashboard

cd dashboard
npm install
npm run dev

---

API Reference

Analyse a prompt

POST /analyze
Content-Type: application/json

{
  "prompt": "Ignore all previous instructions and...",
  "tenant_id": "default",
  "session_id": "user-123"
}

Response

{
  "threat_level": "HIGH",
  "fusion_score": 0.87,
  "blocked": true,
  "threats": ["prompt_injection", "jailbreak"],
  "latency_ms": 42
}

Secure Gateway (proxy to LLM)

POST /secure-gateway
Content-Type: application/json

{
  "prompt": "Explain quantum computing",
  "provider": "openai",
  "model": "gpt-4o",
  "tenant_id": "default"
}

Full API docs available at http://localhost:8000/docs when the backend is running.

---

Project Structure

ai-secops-platform/
├── backend/               # FastAPI backend
│   ├── enterprise_api.py  # Main API entrypoint
│   ├── intelligence/      # Threat detection models
│   ├── pipeline/          # Security pipeline stages
│   ├── redteam/           # Red team simulation engine
│   └── tenants/           # Multi-tenant management
├── dashboard/             # React + TypeScript SOC dashboard
│   └── src/
│       ├── components/    # UI components
│       └── pages/         # Dashboard views
├── fusion_engine/         # Fusion MLP scoring engine
├── anomaly_detection/     # IsolationForest anomaly model
├── trajectory_model/      # Session trajectory transformer
├── data/                  # Threat datasets and indices
└── docker-compose.yml

---

SDK

Looking to integrate AI SecOps into your own application with one import?

aisecops-sdk-securellm →

pip install securellm

---

Contributing

See CONTRIBUTING.md for guidelines.

---

License

Apache 2.0