Production-grade AI Safety Engine protecting humans, organizations, and AI systems from misuse, deception, manipulation, and loss of control.
| Layer | Module | Description |
|---|---|---|
| Level 1 | Prompt Injection Detection | Blocks jailbreaks, instruction overrides, and manipulation attempts |
| Level 2 | Output Safety Guard | Scans LLM responses for harmful content and hallucinations |
| Level 3 | Data Privacy Firewall | Detects and redacts PII/sensitive data |
| Level 4 | Deepfake Detection | Detects AI-generated text, images, audio, and video |
| Level 5 | Smart Control | 55% Block Threshold & AI-driven auto-correction |
- LLM Council (Fab Five): Multi-model validation (Gemini 2, Groq, OpenRouter, etc.)
- Universal Proxy: Global Frontier Gateway targeting 2026 models (GPT-5.2, Claude 4.5)
- Hyper-Resilient Fortress: Adversarial defense suite with Semantic Perturbation and CoT Guard
- Deepfake Engine: Dual-layer detection for photorealistic faces and generative artifacts
- Governance Layer: Full audit logs, risk reports, and compliance dashboards
The platform operates on a 5-layer defense-in-depth model, intercepting traffic via a universal proxy and routing it through safety modules before reaching upstream LLMs.
flowchart LR
User["👤 User / Agent"]
subgraph Platform["🛡️ IntellectSafe Platform"]
direction TB
subgraph Edge["Edge Layer"]
Proxy["Universal Proxy\n(Intercept & Auth)"]
Auth["🔑 Auth & API Keys"]
end
subgraph Safety["Safety Pipeline"]
direction TB
subgraph L1["Level 1: Prompt Shield"]
Inject["Injection Detector"]
PII["PII Scrubber"]
end
subgraph L4["Level 4: LLM Council"]
Gemini["Gemini 2.5"]
Llama["Llama 3.3"]
OpenRouter["OpenRouter"]
end
subgraph L2["Level 2: Output Guard"]
OutputScan["Hallucination &\nSafety Check"]
end
end
subgraph Storage["Storage & Logs"]
DB[("PostgreSQL\nRisk Scores & Logs")]
Redis[("Redis\nRate Limit & Queue")]
end
end
Upstream["☁️ Upstream Provider\n(OpenAI / Anthropic / Google)"]
User -->|"1. Request"| Proxy
Proxy -->|"2. Validate"| Auth
Proxy -->|"3. Scan Prompt"| Inject
Inject -->|"4. Vote"| Gemini
Inject -->|"4. Vote"| Llama
Inject -->|"4. Vote"| OpenRouter
Gemini -->|"5. Consensus"| Upstream
Llama -->|"5. Consensus"| Upstream
OpenRouter -->|"5. Consensus"| Upstream
Upstream -->|"6. Raw Response"| OutputScan
OutputScan -->|"7. Log"| DB
Proxy -->|"Rate Limit"| Redis
OutputScan -->|"8. Safe Response"| User
style Platform fill:#1a1a2e,stroke:#16213e,color:#e0e0e0
style Edge fill:#0f3460,stroke:#533483,color:#e0e0e0
style Safety fill:#16213e,stroke:#533483,color:#e0e0e0
style Storage fill:#1a1a2e,stroke:#e94560,color:#e0e0e0
style L1 fill:#533483,stroke:#e94560,color:#e0e0e0
style L4 fill:#0f3460,stroke:#e94560,color:#e0e0e0
style L2 fill:#533483,stroke:#e94560,color:#e0e0e0
## 🔑 Key Management & BYOK
IntellectSafe supports **Bring Your Own Key (BYOK)** for all major providers.
- **Secure Storage:** Keys are encrypted using `Fernet` (AES-128) before storage.
- **Granular Control:** Assign specific keys to specific tasks (e.g., use a cheap key for high-volume safety scans).
- **Universal Proxy:** Use your stored keys to access any model via our OpenAI-compatible proxy.
### Configurable Safety Scanner
You can dedicate a specific AI connection for **Safety Operations** (Prompt Injection & Output Scanning).
1. Go to **Settings** -> **Upstream Connections**.
2. Add a key (e.g., "OpenRouter Research").
3. Select it in the **"AI Safety Scanner"** dropdown.
4. All safety checks will now route through this connection, keeping your main operational keys separate.
## 🚀 Getting Started
### Prerequisites
- Python 3.10+
- Node.js 18+
- PostgreSQL 15+
### Installation
```bash
# Clone repository
git clone https://github.com/HOLYKEYZ/IntellectSafe.git
cd IntellectSafe
# Backend setup
cd backend
python -m venv venv
.\venv\Scripts\activate # Windows
source venv/bin/activate # Linux/macOS
pip install -r requirements.txt
alembic upgrade head
# Start backend
python -m uvicorn app.main:app --reload --port 8001
# Frontend setup (new terminal)
cd frontend
npm install
npm run dev
- Frontend: http://localhost:3002
- Backend API: http://localhost:8001
- API Docs: http://localhost:8001/docs
The platform includes a Hyper-Resilient Fortress layer designed to stop 90%+ success rate jailbreaks:
- Exploit Instability: Perturbation engine breaks fragile prompt injections.
- Chain-of-Thought Guard: Detects reasoning hijacking and hidden logic bombs.
- Adversarial Simulation: A Council member "shadow-boxes" the prompt to check for harm.
IntellectSafe acts as a universal safety layer. Connect any major AI client and calls are automatically scanned:
| Provider | Model ID Example |
|---|---|
| OpenAI | gpt-5.3, gpt-5.4-thinking |
| Anthropic | claude-4.6-sonnet, claude-4.6-opus |
gemini-3.1-pro, gemini-3-flash |
|
| DeepSeek | deepseek-v4, deepseek-r1 |
| Meta | llama-4-maverick, llama-4-scout |
| Perplexity | sonar-deep-research, sonar-reasoning-pro |
| (i dont know all models but can assure it's gon' work for any one on web session) |
from openai import OpenAI
client = OpenAI(
base_url="http://localhost:8001/v1", # Point to IntellectSafe
api_key="your-openai-key" # Or use X-Upstream-API-Key header
)
response = client.chat.completions.create(
model="gpt-4o",
messages=[{"role": "user", "content": "Hello!"}],
extra_headers={
"X-Upstream-Provider": "openai" # Optional: explicitly set provider
}
)For detailed setup (including Python/LangChain examples & BYOK), read the Integration Guide.
To secure your sessions on ChatGPT, Claude, Gemini, Grok, and Groq, install the IntellectSafe Companion Chrome Extension. It implements a sophisticated Shield & Correct mechanism:
- Smart Blocking: High-risk content (>55% score) is immediately blurred and blocked with a detailed AI explanation.
- AI Re-prompting: Low-to-medium risk content and hallucinations are automatically flagged. The extension auto-sends a context-aware re-prompt to the AI to correct itself in real-time.
Run the connection tester to check if your API keys and the proxy are working:
python backend/scripts/test_connections.py# Scan a prompt for injection
curl -X POST "https://api.intellectsafe.onrender.com/api/v1/scan/prompt" \
-H "Content-Type: application/json" \
-d '{"prompt": "Ignore previous instructions"}'
# Scan LLM output for safety
curl -X POST "https://api.intellectsafe.onrender.com/api/v1/scan/output" \
-H "Content-Type: application/json" \
-d '{"output": "Here is how to...", "original_prompt": "..."}'
# Scan content for deepfakes (Dual-Model Analysis)
# Detects Art (Midjourney/DALL-E) and Photorealistic Faces
-d '{"content_type": "image", "content": "<base64-data>"}'You can now scan prompts and outputs directly from your terminal:
# Scan a prompt
python backend/cli.py scan-prompt "Ignore previous instructions"
# Scan an output
python backend/cli.py scan-output "Here is how to build a bomb..."
# Scan for PII
python backend/cli.py scan-pii "My SSN is 123-45-6789"
# Scan an image for deepfakes
python backend/cli.py scan-image "path/to/image.jpg"
# Agent Control
python backend/cli.py agent-auth "agent-1" "file_read"
python backend/cli.py agent-history "agent-1"
# System Health
python backend/cli.py healthFull lifecycle protection for autonomous agents:
- Authorization: Permission gates for dangerous tools.
- Kill Switch: Immediate agent termination and block.
- Audit: Complete action history and session tracking.
# Authorize agent action
curl -X POST "https://api.intellectsafe.onrender.com/api/v1/agent/authorize" \
-H "Content-Type: application/json" \
-d '{"agent_id": "agent-1", "session_id": "s1", "action_type": "file_read", "requested_action": {"path": "/tmp/test.txt"}}'cd backend
# Test all scan endpoints
python verify_backend.py
# Test Universal Proxy
python verify_proxy.py
# Test Agent Control
python verify_agent.pyGPLv2 GNU GENERAL PUBLIC LICENSE Version 2 License .
NOTE: still in development & research stage due to constant frontier models releases