An automated runtime testing and security monitoring middleware designed to secure LLM-based AI Agents against prompt injections, alignment drift, and critical logic failures.
As autonomous AI Agents take on execution capabilities in decentralized applications and Web3 environments, they become vulnerable to external manipulation. AgentWatchdog AI acts as an inline proxy and continuous evaluation subsystem that intercept inputs/outputs, neutralizing attacks before they compromise application state or breach data integrity.
- Real-Time Guardrails: Pre-scans incoming user requests and prompts to detect and neutralize adversarial injection payloads.
- Continuous Turing Benchmarking: Evaluates AI Agent health during idle states via automated synthetic challenges to measure logic drift.
- Immutable Audit Trail: Pre-architected to push core cryptographic integrity verifications and security logs onto the Mantle Network L2 chain.
[ User / Client ] │ ▼ ┌──────────────────────────────────────┐ │ AgentWatchdog Proxy │ ◄── [ Threat Inspection & Guards ] └──────────────────┬───────────────────┘ │ (Filtered Prompt) ▼ ┌──────────────────────────────────────┐ │ Target AI Agent │ └──────────────────┬───────────────────┘ │ (Response Data) ▼ ┌──────────────────────────────────────┐ │ Mantle L2 On-Chain Log │ ─── [ Immutable Security Audit ] └──────────────────────────────────────┘ 💻 Tech Stack • Backend Framework: Node.js (TypeScript) / Python (FastAPI) • AI Security Integration: Advanced System Prompt Hardening & Token Scanners • Blockchain Ecosystem: Mantle Network (L2 Testnet Core for immutable logs) 🚀 Getting Started (Implementation files and environment setup instructions coming soon as development progresses during the Turing Test Hackathon).