Early access for legal and regulated teams now open Request access →

Protection · Real-time Scanning

Every prompt. Scanned before it sends.

Ventrin's on-device NLP model runs in under 50 milliseconds — before the browser fires a single byte. By the time the AI tool receives the request, every identifier has already been assessed and handled.

<50ms
Detection latency — measured from the moment a user hits send to the point the prompt is cleared or redacted
97%
PII recall rate on legal document test sets — names, IDs, dates and organisation references caught before transmission
0
Remote calls made during detection — the model runs entirely in the browser sandbox with no network dependency

How it works

Intercept. Classify. Clear.

01

The extension hooks the send event.

Ventrin injects a lightweight content script into supported AI tools — ChatGPT, Claude, Gemini and Copilot. When a user submits a prompt, the extension intercepts the form submission or API call before the browser network layer fires. The user sees nothing different.

02

An ONNX model classifies tokens locally.

A compact Named Entity Recognition model — bundled with the extension at install time — runs over the prompt text. It identifies person names, organisation names, dates of birth, ID numbers, API keys and other sensitive token types. The entire classification happens inside Chrome's extension sandbox with no outbound network calls.

03

Policy determines the outcome.

Based on the rules your compliance team has configured, flagged tokens are either replaced with labelled placeholders like [NAME] and [DATE_OF_BIRTH], or the send is blocked with a clear explanation. Every event is logged in the admin dashboard regardless of outcome.

Why it matters

Built for the prompt, not legacy DLP.

Traditional data loss prevention tools were designed for files and email. They add hundreds of milliseconds of latency, require cloud connectivity, and were trained on document-length text. Ventrin's model was purpose-built for the AI prompt: short, conversational, high-entropy — the format your team actually uses every day.

  • Detects PII in mixed-format prompts — mid-sentence, abbreviated, or comma-separated
  • Updates silently with each extension release — no re-deployment by IT
  • Zero latency penalty for prompts that contain no sensitive data
  • False-positive rate tuned for legal context — firm names and case references handled correctly
Scan result · example
Latency 42 ms
Tokens scanned 84
Client name REDACTED
Date of birth REDACTED
Organisation ref FLAGGED
Policy outcome SENT CLEAN

Scan pipeline · timing breakdown

⌨️

User submits

t = 0 ms

🪝

Send intercepted

+1 ms

🧠

NER model runs

+8–48 ms

Policy applied

+1 ms

Measured on a 2021 MacBook Pro M1 · 84-token prompt · p95 latency. Zero network calls made during scan.

How Ventrin compares to legacy DLP

Capability Ventrin Real-time Scan Traditional DLP (e.g. Symantec, Forcepoint) Browser content policy
Scans AI prompt text Yes Partial — file/email only No
On-device — no cloud relay Yes No — cloud proxy required Yes
Latency added to prompt < 50 ms 200–800 ms typical 0 ms (no scan)
Legal-context NER (firm names, matter IDs) Yes No No
API key / credential detection Yes — 20+ formats Limited regex only No
Per-team policy granularity Yes Enterprise tier only No
Audit log per-prompt Yes File/email events only No
Zero-install for end user Chrome extension Agent required on device Browser built-in

Protection that runs at the speed of thought.

Install in two minutes. Scanning starts immediately.

Manifest V3 · Local detection · No prompt storage by default.