Powered by AMD MI300X · Trained on 1,000,000 Real-Chain Graphs · EIP-8259 Native
Sigui is a decentralized security infrastructure (DePIN) for AI agents, utilizing the massive parallel power of AMD MI300X GPUs to detect on-chain threats via visual graph analysis.
In < 50ms, the 5-layer Risk Engine, Imina Na V2 vision model (trained on 1,000,000 samples), and MemoClaw memory return a verdict: ALLOW / BLOCK / ESCALATE.
Blocked attacks are permanently recorded in ThreatRegistry.vy on Arc L1.
Risk thresholds are governed by the community via Hogonat.vy staking.
The loop is fully closed. Zero human intervention.
Sigui now includes a revolutionary Agent DID (Decentralized Identity) system that solves the fundamental problem of agent identity verification:
The Problem: Currently, any attacker can create a new wallet, build fake reputation for 2 weeks, then execute attacks. Agent identity is tied to wallets, not cryptographic identity.
The Solution: Agent DID system where reputation is cryptographically bound to identity, not wallets.
- Cryptographic Identity: Each agent has unique DID bound to Ed25519 public key
- Portable Reputation: Reputation follows identity across wallets/chains
- Impossible to Usurp: Cannot fake identity without private key
- Reputation Laundering Prevention: Cannot "wash" bad reputation by creating new wallets
- Multi-Tier Verification: Bronze → Silver → Gold → Platinum verification levels
| Tier | Requirements | Trust Multiplier | Benefits |
|---|---|---|---|
| Bronze | DID registration + email | 1.2x | Basic reputation |
| Silver | Organization verification | 1.5x | Enhanced scoring |
| Gold | Enterprise audit + KYC | 1.8x | Premium features |
| Platinum | Insurance backing | 2.0x | Maximum trust |
- AMD MI300X Optimized: 10-100x performance improvement over CPU
- 6-Factor Scoring: Identity (25%) + Transactions (30%) + Verification (20%) + Cross-chain (10%) + Threat Intel (10%) + Insurance (5%)
- Real-time Updates: Reputation updates in <50ms
- Anomaly Detection: Isolation Forest for detecting suspicious behavior
- Predictive Scoring: ML model trained on 10,000+ agent behaviors
did:sigui:arc:individual:0x7a3f2e1b9c8d4e5f
Components:
did: Standard DID prefixsigui: Sigui method identifierarc: Blockchain network (Arc L1)individual: Agent type (individual/organization/enterprise)0x7a3f...: Public key hash (16 characters)
AI Agent ──── POST /evaluate ($0.001 x402) ────────────────────────────┐
│
┌───────────────────────────────────────────────────────────┐ │
│ Sigui Security Pipeline │ │
│ │ │
│ ① Agent DID Check (cryptographic identity) │ │
│ ② Behavior Engine (numpy, < 2ms) │ │
│ ③ Anti-Splitting (flow tracking, cumulative USDC) │ │
│ ④ Service Registry (reputation scores, on-chain) │ │
│ ⑤ Contract Inspector (bytecode analysis, approvals) │ │
│ ⑥ Imina Na Vision (AMD MI300X ROCm, < 50ms) │ │
│ │ │
│ PolicyBrain ←→ LangGraph ←→ CrewAI (self-critique) │ │
└───────────────────────────────────────────────────────────┘ │
│
ALLOW ──────────────────────────────────────────────────────-─┘
BLOCK ──── ThreatRegistry.vy (Arc L1, permanent, immutable)
ESCALATE ── Qwen2.5 (AMD MI300X) → Claude 3.5 fallback
Agent pays $0.001 ──→ Sigui Oracle receives revenue
│
┌───────┴────────┐
80% 20%
│ │
Treasury Hogonat.vy
(pays Claude & Fee Pool
Arc fees) │
Stakers claim
proportional rewards
│
Stakers vote
risk_weights &
thresholds
│
PolicyBrain
reloads live thresholds
│
←── Better protection → more agents
Every evaluation strengthens the protocol. Every staker earns from every attack blocked.
Traditional security engines use simple thresholds. Imina Na sees the attack topology.
- Architecture: Fine-tuned Qwen2-VL-2B-Instruct (LoRA)
- Hardware: AMD MI300X (ROCm stack) for training and inference
- Dataset: The Dogon Dataset — 10,000+ custom transaction graph topologies
- Patterns detected:
DRAIN_STAR,MIXING_CHAIN,COORDINATED_CLUSTER - Integration: Provides a
risk_deltainjected into the Kanaga risk aggregator
| Component | Technology | Role |
|---|---|---|
| API | FastAPI + uvicorn | Security oracle endpoint |
| Payment | x402 protocol + Circle DCW | $0.001/call micropayments |
| Risk Engine | numpy, 5-layer pipeline | Real-time threat scoring |
| Vision Oracle | Imina Na (Qwen2-VL, LoRA) | Graph topology detection |
| AI Hardware | AMD MI300X (ROCm) | Training + inference |
| Memory | MemoClaw (SQLite async) | Episodic memory, pattern recall |
| Self-Critique | PolicyBrain + LangGraph + CrewAI | Autonomous threshold tuning |
| Smart Contracts | Vyper 0.4.3 (2 contracts) | ThreatRegistry + Hogonat DAO |
| Blockchain | Arc L1 Testnet | Immutable attack log |
| Governance | Hogonat.vy staking | Community votes on risk thresholds |
| Dashboard | Next.js 14 | Cinematic live monitoring UI |
| SDK | Python (pip install sigui) | 2-line integration for agents |
# Python 3.11+ required
pip install -r requirements.txt
# For AMD GPU inference (ROCm):
pip install -r requirements_amd.txtcp .env.example .env
# Edit .env with your Circle API key and Arc testnet credentials
# DEMO_MODE=true works out of the box with no credentialsuvicorn main:app --port 8000 --reload
# API: http://localhost:8000
# Swagger: http://localhost:8000/docscd demo-ui
npm install && npm run dev
# Dashboard: http://localhost:3001Click ⚡ Lancer le rituel in the dashboard to launch the 5-agent simulation:
- 🔥 Danseur du Feu (PayerAgent) — legitimate payments, varied amounts
- 🦊 Renard Pâle (AttackerAgent) — adversarial, 4 rotating strategies
- ⭐ Étoile Apprenante (LearnerAgent) — warmup → probe → attack cycles
- 🌫 Gray Zone (GrayZoneAgent) — ambiguous transactions challenging the AI
- 👁 Œil de la Société (MonitorAgent) — ROI tracking, anomaly detection
pip install siguifrom sigui import SiguiClient, AgentIdentity
# Create agent identity
identity = AgentIdentity.create(
agent_type="trading_bot",
verification_tier="silver"
)
async with SiguiClient(api_url="http://localhost:8000") as client:
# Register identity
await client.register_identity(identity)
# Evaluate transaction with identity
result = await client.evaluate_with_identity(
amount=5.0,
destination="0xRecipient",
agent_did=identity.did
)
print(f"✅ ALLOW — reputation={result.reputation_score:.3f} — {result.processing_time_ms}ms")pip install siguifrom sigui import SiguiClient
async with SiguiClient(api_url="http://localhost:8000") as client:
result = await client.evaluate(amount=5.0, destination="0xRecipient")
if result.is_safe:
print(f"✅ ALLOW — risk={result.risk_score:.3f} — {result.processing_time_ms}ms")
else:
print(f"🚫 BLOCK — {result.reason}")| Metric | Value |
|---|---|
| ThreatRegistry | 0x17430A67e11535466cC5f17e736D5e4643B86ba1 |
| Hogonat DAO | 0x1e497ed89fc4f6ae2Da826a18B4a7728919684a1 |
| Explorer | testnet.arcscan.app |
| Attacks recorded | 380+ blocked and permanently logged |
| Model | Ibonon/imina_na_lora |
| Feature | Traditional Security | Sigui |
|---|---|---|
| Response | Passive filter | Autonomous economic agent |
| Identity | Wallet-based | Cryptographic DID |
| Reputation | Per-wallet | Portable across chains |
| Learning | Static rules | MemoClaw episodic memory + self-critique |
| Governance | Centralized | On-chain staking DAO (Hogonat.vy) |
| Payments | Subscription | x402 pay-per-call ($0.001) |
| Proof | Logs | Immutable on-chain (ThreatRegistry.vy) |
| Vision | None | Imina Na graph topology (AMD MI300X) |
| Audit trail | Centralized DB | Arc L1 blockchain |
docker-compose up
# Backend: http://localhost:8000
# Dashboard: http://localhost:3001Sigui/
├── main.py # FastAPI app — 10-step startup sequence
├── modules/
│ ├── gateway.py # All routes, x402 middleware (50KB)
│ ├── security_engine.py # 5-layer risk engine (numpy)
│ ├── memory.py # MemoClaw episodic memory (SQLite async)
│ ├── ai_engines.py # PolicyBrain + CrewAI + LangGraph + Claude
│ ├── imina_na_vision.py # Imina Na vLLM client (AMD MI300X)
│ ├── response_validator.py # LLM response poisoning detection
│ ├── treasury.py # Autonomous treasury management
│ └── identity/ # 🆕 Agent DID system
│ ├── agent_did.py # DID generation & management
│ ├── reputation_engine.py # AI-powered reputation scoring
│ └── identity_integration.py # Identity + security pipeline
├── contracts/
│ ├── ThreatRegistry.vy # Onchain attack log (Vyper 0.4.3)
│ ├── Hogonat.vy # Governance DAO — staking + threshold votes
│ └── AgentIdentityRegistry.vy # 🆕 Agent identity registry
├── ecosystem/
│ └── agents.py # 5 autonomous agents (Payer, Attacker, Learner...)
├── governance/
│ └── hogonat_client.py # Dual-mode governance client (mock/on-chain)
├── sdk/python/sigui/ # Python SDK (pip install sigui)
└── demo-ui/ # Next.js 14 live dashboard
Phase 1 ✅ Agent Identity System - Cryptographic identity for AI agents
Phase 2 🔄 Threat Intelligence Network - Decentralized marketplace for attack patterns
Phase 3 🔄 Insurance Layer - Risk coverage for autonomous transactions
Phase 4 🔄 Open Protocol Standard - EIP-XXXX Agent Security Standard
The Goal: By 2030, every AI agent that moves value goes through Sigui. We're building the trust infrastructure for the autonomous economy.
Sigui · Built for the Agentic Economy · Arc + Circle Hackathon · 2026
"The firewall doesn't wait for humans. It learns, adapts, and protects — autonomously."
Next Steps:
- Deploy AgentIdentityRegistry.vy to Arc mainnet
- Onboard 1,000 agents with cryptographic identities
- Launch Threat Intelligence Network (Phase 2)
- Submit EIP-XXXX Agent Security Standard (Phase 4)