🛡️ SentinelAISecurity Analyst Portal

SentinelAI — Install & Usage

Real-time DLP for AI tools. Three on-device pieces: a browser extension (sensor), a local agent (rules + risk + optional on-device AI), and Ollama (the small local model). Nothing sensitive leaves the laptop; only event records go to the central dashboard.

Live dashboard: https://dlp.kshetra.studio

Windows (for the sponsor)

  1. Unzip SentinelAI-Windows.zip.
  2. In PowerShell, from the unzipped folder:
    .\install-windows.ps1
    This installs Ollama + the model, sets up the agent, and registers it as a Scheduled Task (auto-starts at logon, restarts on crash). Requires Python 3.11–3.13 (winget install Python.Python.3.12 if needed).
  3. Load the extension: edge://extensions or chrome://extensions → Developer mode → Load unpacked → select the extension folder from the zip.
  4. Confirm the SentinelAI popup shows a green dot ("agent connected").

Manage it: Task Scheduler → "SentinelAI Agent", or .\uninstall-windows.ps1 to remove.

macOS

./packaging/install-macos.sh

Installs Ollama + model, sets up the agent, registers launchd services (auto-start + restart), then load the extension via chrome://extensions → Load unpacked → extension/.

Try it (talking points for a demo)

With the extension loaded and the popup dot green, open ChatGPT/Claude and send:

Prompt What happens
my ssn is 123-45-6789 Blocked (PII)
card 4111 1111 1111 1111 Blocked (PII, Luhn-checked)
import os + a function using it Blocked (source code)
A public snippet (e.g. quicksort) with AI toggle ON Allowed — the on-device AI suppresses the false positive
Code containing orion (a proprietary marker) with AI ON Blocked — org policy overrides the AI

Every decision appears on the dashboard with a Signal column (Rules / AI-suppressed / Fail-safe) and analysts can mark decisions correct/wrong (feedback loop).

Troubleshooting

  • Nothing blocks: check the popup dot. Red = agent not reachable. Windows: is the "SentinelAI Agent" task running? macOS: launchctl list | grep sentinelai.
  • A specific secret "passes": password managers can mask an API-key-looking string into bullets before the page sees it. Test with an SSN / card / code snippet instead of a raw sk-… key.
  • AI toggle does nothing: Ollama must be running with the model pulled (ollama pull llama3.2:3b). If it's absent, the agent safely falls back to rules-only.

What runs where

  • On the laptop: extension, agent (rules + AI decision), Ollama (the model). Prompt text never leaves the device.
  • In the cloud (your AWS): only the event records (metadata + a truncated snippet) that power the dashboard.