Sentinel

AI agent security platform, install once, protect everything

Developer Tools

Artificial Intelligence

GitHub

Security

Hey Product Hunt! 👋 I'm Partha. My co-founders Kaushik and Junaid and I built Sentinel because we kept running into the same problem: teams ship AI agents fast, but security is always bolted on after something goes wrong. Some of you may remember our earlier launch — Breach Intel — which focused on compliance breach detection. It answered "what did my AI agent do wrong?" with an immutable audit log. But after deploying it, we realized detection alone isn't enough. Teams kept asking: "Can it also block the bad stuff before it happens?" So we built Sentinel — a unified platform that wraps Breach Intel with seven additional security layers. One install script. Zero code changes. Every Python agent using OpenAI, Anthropic, or LangChain is automatically protected. **What it prevents:** 🔒 Prompt injection — 3-layer detection (<1ms pattern matching, <50ms risk scoring, local LLM for edge cases) 🛡️ PII & credential leaks — 9 DLP scanners block SSNs, credit cards, API keys before they leave 🍯 Honeypot traps — decoy .env, .aws/credentials, .ssh/id_rsa that alert on access 🛑 Block rules engine — keyword, regex, recipient, and channel rules with seeded defaults **What it detects:** 📋 40+ compliance breach types across Fintech (PCI-DSS, SOX), Healthcare (HIPAA), Pharma (FDA 21-CFR) ⚡ Rule-based classifier — deterministic, <1ms, no LLM in the critical path 🔗 SHA-256 immutable audit chain — no UPDATE, no DELETE, tamper detection built in **What it shows you:** 📊 Unified dashboard — 6 tabs: Overview, Live Events, Traces, Breach Monitor, Block Rules, Settings 📡 LLM proxy tracks cost, latency, and tokens across 16 models 🔍 Expandable trace visualization with per-span cost and latency 💬 Channel security — blocks sensitive data from reaching Telegram, WhatsApp, Discord, Slack 🐳 Hardened containers — non-root, read-only rootfs, custom seccomp profiles **How it works:** 1. Run ./install.sh 2. Four services auto-start on boot (Policy Engine, Breach Intel, LLM Proxy, OpenClaw Gateway) 3. sitecustomize.py patches all LLM frameworks at import time — zero code changes 4. 8 security hooks inspect every stage: message received → tool call → prompt build → LLM input → LLM output → message sending 5. Open localhost:8000/unified to see everything **The difference from Breach Intel:** Breach Intel = 1 layer (compliance logging) → "what violations did my AI commit?" Sentinel = 8 layers (prompt blocking + DLP + endpoint monitoring + LLM proxy + channel security + honeypots + containers + compliance) → prevents them AND logs them. **Key numbers:** - 8 security hooks, 5 endpoint monitors, 9 DLP scanners - 40+ breach types across 3 regulated verticals - <1ms breach classification, 3-layer prompt analysis - 74 Python modules, unified from 5 repositories We'd love your feedback — especially if you're running AI agents in regulated industries. What security gaps are you seeing? Happy to answer any questions! 🙏 🔗 Website: https://parthamehtaorg.github.io... 📖 Docs: https://parthamehtaorg.github.io... 📦 GitHub: https://github.com/ParthaMehtaOr... Partha (@partha_mehta), Kaushik (@kaushik_dharamshi), Junaid (@junaid_mundichipparakkal)