🛡️ SentinelAISecurity Analyst Portal

Stop secrets leaking into AI tools — in real time, on the device.

SentinelAI intercepts what your team types into ChatGPT, Claude, Gemini and Copilot before it is sent, classifies it locally, and allows, warns, or blocks. Nothing sensitive leaves the laptop.

The problem

Employees paste API keys, customer PII, source code, and unreleased plans into AI chats. It is invisible to security, and once it is sent, it is gone. Network DLP can’t see inside TLS to these apps, and after-the-fact detection is too late.

Three pieces, all on the endpoint

Browser extension

Hooks the request the moment a prompt is sent and holds it for a verdict.

Local agent

Deterministic rules + risk scoring + policy. Decides allow / warn / block in <5 ms.

On-device model

A small local LLM (via Ollama) adds judgment — optional, off by default.

Prompt → extensionagent (rules + AI) → allow / warn / block → event to dashboard

Two layers of protection

1 · Rules (always on)

Fast, deterministic detection of structured data: API keys (OpenAI, AWS, Slack, Stripe…), PII (SSN, credit cards with Luhn/IBAN checks), source code, financial & contract language. Hard secrets always block.

2 · AI (optional toggle)

An on-device model works in both directions — it removes false alarms and catches leaks the rules can’t pattern-match — tailored to your organization.

The AI works both ways

InputRules onlyRules + AI
Suppress — a public code sample (quicksort)blockallow
AI-suppressed: public sorting algorithm
Detect — “shutting the Mumbai office… don’t share externally”allow
rules see nothing
block
AI-detected: internal relocation plans
Control — “what is the capital of France?”allowallow

Safety by design: the AI can only relax fuzzy categories and only raise allows — at worst it over-blocks, it can never cause a leak. Real secrets always block via rules.

Tailored to your organization

A per-org / per-department policy teaches the model your context — company description, proprietary codenames, and “public is fine” hints. Anything carrying a proprietary marker (e.g. a project codename) is never suppressed. Edit the policy and it takes effect live.