AgentWatch — AI Agent Security Platform

AgentWatch is a real-time AI agent security platform built on Veea Lobster Trap, securing both Google Gemini and Anthropic Claude backends equally. THE PROBLEM: AI agents read files, hit APIs, and trigger actions in production — but guardrails haven't kept up. A single malicious prompt can leak credentials or exfiltrate data. Most teams have zero visibility into what their agents are doing. WHAT WE BUILT: Deep Prompt Inspection: Every prompt analyzed for injection attempts, jailbreak signatures, PII, credential leakage, and exfiltration patterns via Lobster Trap DPI. Risk scores 0.0–1.0 in real-time. Multi-LLM Security: AgentWatch secures both Gemini 2.0 Flash, Gemini 2.0 Pro, and Claude Sonnet. The same Lobster Trap security layer applies to all backends equally — switch models without losing governance. Policy Enforcement: First-match-wins YAML rules compatible with Lobster Trap. Actions: ALLOW, DENY, LOG, HUMAN_REVIEW, QUARANTINE, RATE_LIMIT. HIPAA, SOC2, and finance packs included. Real-Time Dashboard: Blazor WebAssembly plus SignalR — every attack and alert appears live. Risk trend charts, threat breakdowns, live event feed. Full Audit Trail: Every interaction logged with declared vs detected intent, risk score, and threat flags. CSV export for regulators. LIVE DEMO RESULTS: - Prompt Injection: 0.95 risk → BLOCKED - Data Exfiltration: 0.98 risk → BLOCKED - PII Detected: 0.55 risk → LOGGED - Intent Mismatch: 0.78 risk → HUMAN_REVIEW - Clean Request: 0.02 risk → ALLOWED STACK: ASP.NET Core 10, Blazor WebAssembly, SignalR, SQLite, Google Gemini API, Anthropic Claude, Veea Lobster Trap. Lobster Trap is the floor. AgentWatch is the ceiling.