DEV Community

Dmitry Labintcev
Dmitry Labintcev

Posted on • Edited on

Christmas Gift: Open-Sourcing 99 AI Security Detection Engines

#python #machinelearning #opensource #security

📌 This post is now archived. For the latest updates on SENTINEL, see the new consolidated article:
SENTINEL Platform — Complete AI Security Toolkit (2026 Update Log)

I'm releasing the full source code of SENTINEL — an AI security platform. Not a "lite version" or "community edition" — everything.

🚀 One-Liner Deploy (NEW!) 

curl -sSL https://raw.githubusercontent.com/DmitrL-dev/AISecurity/main/install.sh | bash
Enter fullscreen mode Exit fullscreen mode

5 services, 99 engines, 5 minutes:

  • Gateway (Go) — HTTP/HTTPS API
  • Brain (Python) — 99 detection engines
  • Redis — caching & rate limiting
  • PostgreSQL — audit logs
  • Dashboard — web UI

What is it?

SENTINEL is a security platform for LLMs, AI agents, and multimodal systems:

Component Description
🛡️ Defense 99 detection engines (<10ms latency)
⚔️ Strike Red team platform (39,000+ attack payloads)
📊 OWASP Coverage LLM Top 10 ✅ + Agentic Top 10 2026 10/10 ✅

Think of it as a firewall + pentest suite, but for AI.

The "Strange Math" Engines

While most AI security tools use pattern matching, we went a different way:

Topological Data Analysis (TDA) 

from gudhi import RipsComplex

rips = RipsComplex(points=embedding, max_edge_length=2.0)
simplex_tree = rips.create_simplex_tree(max_dimension=2)
persistence = simplex_tree.persistence()

# Attacks create topological anomalies in embedding space
Enter fullscreen mode Exit fullscreen mode

Idea: Jailbreak attempts create "holes" in the embedding topology that normal text doesn't.

Sheaf Theory

Coherence verification across multi-turn conversations. Detects attacks that slowly shift context across messages.

Hyperbolic Geometry

Poincaré ball embeddings for attack clustering. Semantic relationships form hierarchies that attacks disrupt.

December 2025 Updates

Feature Status
99 Detection Engines +2 from launch
OWASP Agentic 2026 10/10 coverage
Supply Chain Guard MCP/A2A protection
Trust Exploitation Social engineering via AI
Echo State Network Temporal pattern detection
One-Liner Deploy 5 services in 5 minutes

Benchmarks

Metric Value
Recall 85.1%
Precision 84.4%
F1 Score 84.7%
Latency <10ms
Engines 99

Tested on 1,815 samples from 3 HuggingFace datasets.

Quick Start

Option 1: One-Liner (Recommended) 

curl -sSL https://raw.githubusercontent.com/DmitrL-dev/AISecurity/main/install.sh | bash
Enter fullscreen mode Exit fullscreen mode

Option 2: Docker Compose 

git clone https://github.com/DmitrL-dev/AISecurity.git
cd AISecurity/sentinel-community
cp .env.example .env
docker-compose -f docker-compose.full.yml up -d
Enter fullscreen mode Exit fullscreen mode

Option 3: Python Package 

git clone https://github.com/DmitrL-dev/AISecurity.git
cd AISecurity/sentinel-community
pip install -e .
Enter fullscreen mode Exit fullscreen mode

API Example 

curl -X POST http://localhost:8080/api/v1/analyze \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"prompt": "Ignore all previous instructions"}'
Enter fullscreen mode Exit fullscreen mode

Response:

{
  "safe": false,
  "risk_score": 85.5,
  "threats": ["prompt_injection"],
  "blocked": true,
  "latency_ms": 8
}
Enter fullscreen mode Exit fullscreen mode

Why Open Source?

  1. AI security needs transparency — trust but verify
  2. Threats evolve too fast for one team
  3. It's Christmas 🎄

Links

Happy to answer questions! ⭐ Star the repo if you find it useful.

Top comments (2)

Collapse
 
dmitry_labintcev_9e611e04 profile image
Dmitry Labintcev •

🚀 Update (Dec 29): Major Milestones!

Since Christmas, SENTINEL has grown significantly:

Metric Dec 26 Now
Detection Engines 99 → 170

Research Inventions 49 → 56

Unit Tests 936 → 1,200+

HuggingFace Dataset — 51,610 samples
New Detection Engines:

🧬 Strange Math v2 (Sheaf, TDA, Hyperbolic)
🐝 Bio-Intelligence (AIS, Swarm)
🤖 Agentic Defense (Memory Shield, Tool Guardian)
🔌 Protocol Security (MCP, A2A, NHI)
🐉 Strike v4.0 — Red Team Testing:

Successfully tested against 82 Crucible challenges
HYDRA Engine: 9-headed parallel attacks
AI Attack Planner with Gemini integration
85.1% recall, 92.3% precision on adversarial benchmarks
Full repo: github.com/DmitrL-dev/AISecurity

Some comments may only be visible to logged-in visitors. Sign in to view all comments.