SentinelAI — Secure Enterprise Agent Platform

Enterprises want AI agents. Their security teams say no — because nobody can answer: what did the agent just do, could it have leaked data, and can a regulator audit it? SentinelAI solves all three by combining Gemini-powered specialist agents (HR, Legal, Finance, DevOps) with a deep prompt inspection security layer that sits between every user message and the AI model. Every prompt is inspected bidirectionally for intent, PII, credential leaks, prompt injection, and exfiltration patterns before Gemini processes it — and the response is inspected too. Approved requests get answered using Gemini Pro with RAG over a company knowledge base. Every action — allowed or blocked — is logged with risk score, policy ID, intent category, and timestamp. What makes SentinelAI unique: - Behavioral DNA Profiler — tracks 5 behavioral dimensions per agent (risk, sentiment, formality, certainty, latency) and fires a Z-score drift alert when an agent starts behaving unusually — detecting jailbroken or compromised agents before damage occurs. A completely novel security primitive. - Declared vs Detected Intent — agents declare intent; the DPI layer detects actual behavior. Mismatches scored 0–1 trigger human review automatically. - Red Team Attack Simulator — 10 OWASP LLM Top 10 attack vectors run automatically, generating a Security Score and letter grade for your agent deployment. - Bidirectional DPI — every prompt AND every response is inspected, not just inputs. - 1-Click Compliance Reports — SOC2, HIPAA, and Finance audit reports exportable in seconds. Regulator-ready, timestamped, fully cited. Built on Gemini Pro and Flash via Google AI Studio, with YAML-based policy packs for HIPAA, SOC2, and Finance compliance. Deployed on Streamlit. Fully open source. Tracks: Agent Security & AI Governance + AI Agents with Google AI Studio.