AgentShield: Pre-Execution Security Firewall for Autonomous AI Agents
Firewall pre-ejecución para agentes de IA. Audita transacciones en tiempo real usando políticas on-chain (Somnia/X Layer) y análisis de LLMs antes de firmar. Verdicts: ALLOW, WARN, BLOCK.
视频
技术栈
描述
AgentShield is a pre-execution security firewall for autonomous AI agents built for the OKX AI Genesis Hackathon.
🔐 THE PROBLEM
AI agents are increasingly executing on-chain transactions autonomously. Without a security layer, they can approve unlimited token spending, fall for phishing scams, or drain wallets. AgentShield sits between the agent and the blockchain, analyzing every action before it executes.
⚙️ HOW IT WORKS
1. User defines a security policy in plain language (e.g. "Block scams, max 50 STT per tx")
2. AI agent proposes an on-chain action
3. AgentShield runs deterministic checks (instant, free) — catches obvious attacks
4. If not blocked, Gemini 2.5 Flash performs deep LLM analysis
5. Returns verdict: ALLOW, WARN, or BLOCK with risk score (0-100)
🏗️ ARCHITECTURE
- Smart contracts on Somnia Testnet (7 contracts, 171 tests, 0 failures)
- MCP endpoint powered by Gemini 2.5 Flash for real-time AI analysis
- Registered as ASP #5936 on OKX.AI marketplace
- Frontend demo: agentshield-dusky.vercel.app
🛡️ SECURITY
- Deterministic-first: hard blocks cost zero gas, no LLM needed
- Fail-safe: unexpected AI output defaults to WARN, never ALLOW
- On-chain verifiability: all scans stored permanently
- 171 tests, production-ready, not a prototype
🏆 WHY AGENTSHIELD WINS
- Complete product: contracts + endpoint + demo, all functional
- Real security value: protects agents from the #1 attack vectors
- Dual-chain: Somnia for LLM inference, OKX.AI for distribution
- Best Product + Software Utility + Finance Copilot categories