<RajVardhan/>
Back to catalog
Cybersecurity
2026-09-25

VEIL-X — Privacy-First Visual Intelligence for Browser Agents

On-device visual perception layer for browser agents that redacts passwords, Aadhaar/PAN, cards, and other PII inside the browser, so raw screenshots never leave the device.

VEIL-X — Privacy-First Visual Intelligence for Browser Agents Preview

The Problem

Browser agents usually send raw screenshots and full page DOM to cloud vision-language models. In banking, government, or defense apps, that leaks passwords, identity numbers, and other sensitive data over the internet.

The Solution

VEIL-X runs a zero-leakage perception pipeline: it captures the viewport in memory, detects sensitive DOM elements locally, redacts them on an HTML5 canvas, and runs a small vision model in the browser. Only the sanitized image and local context reach the backend VLM, which returns a structured action plan.

Key Capabilities
  • On-device PII detection for passwords, emails, phone numbers, Aadhaar, card numbers, and secret PINs using token heuristics and regular expressions.
  • In-browser canvas redaction with blackout, Gaussian blur, or scrambled placeholders. The unredacted image never leaves the browser.
  • Client-side scene classification with ONNX Runtime Web (WebGPU / WASM SIMD) using a 4.95 MB SqueezeNet model.
  • FastAPI backend that queries Gemini, OpenAI, or an offline VLM and returns click / type / scroll actions as JSON.
  • Content-script action executor with synthetic pointer and keyboard events and visual action indicators.

Technology Stack

JavaScriptChrome Extension APIsHTML5 CanvasONNX Runtime WebPythonFastAPIGemini / OpenAI APIs

Zero-Leakage Pipeline

  1. Capture the visible viewport in volatile memory only.
  2. Detect sensitive DOM elements and text nodes locally.
  3. Redact them on an HTML5 canvas before anything is transmitted.
  4. Classify the scene in the browser with ONNX Runtime Web.
  5. Reason on the backend using only the sanitized image and local context.
  6. Act in the page through the content script.

Lessons Learned & Key Takeaways

VEIL-X taught me to design AI systems privacy-first: deciding what data must never leave the device, running inference in the browser with WebGPU and WASM, and keeping agent actions safe and explainable.