SENTINEL — AI Agent Governance Layer

SENTINEL sits between AI agents and language model backends as a two-layer governance system. The first layer is Veea Lobster Trap: a deep packet inspection proxy that evaluates every agent prompt against 14 pre-configured policy rules in under one millisecond. It detects and blocks prompt injection, malware requests, phishing, data exfiltration, sensitive file access, PII exposure, dangerous system commands, base64-obfuscated payloads, and credential leaks in model output. Verdict: ALLOW, DENY, or HUMAN_REVIEW. The second layer is Gemini 2.5 Flash: for every request — allowed or blocked — Gemini produces a structured governance report with risk level (LOW / MEDIUM / HIGH), incident summary, reasoning, and recommended operator action. When a prompt passes all policy rules but carries semantic risk — role impersonation, authority escalation, intent manipulation — Gemini escalates it to HUMAN_REVIEW. A second Gemini agent reviews the first agent's analysis before issuing the final governance verdict. The operator dashboard updates in real time. For each captured incident, the operator sees the original agent prompt, the Lobster Trap enforcement verdict and matched rule, SENTINEL's recommended verdict, Gemini's full reasoning, and a workflow decision: Approved, Rejected, or Quarantined. The SENTINEL MCP server exposes 10 governance tools over stdio, callable from Claude (Claude Desktop / Claude Code), GPT-4o (OpenAI Agents SDK), Gemini (Google ADK / MCPToolset), or any local Llama model via Ollama. The same governance workflow is accessible from any enterprise AI stack with no vendor lock-in. Built for Track 1 (Agent Security & AI Governance), Track 2 (AI Agents with Gemini / AI Studio), and Track 4 (Data & Intelligence).