Research Summary: Penny for Your Prompts: A Measurement Study of Contextual Leakage in AI Chat Platforms

A measurement study examining 42 general-purpose and specialized AI chatbot platforms (spanning health, legal, and education domains) demonstrates that standard web tracking technologies—including analytics, advertising, and session replay scripts—operate silently alongside conversational interfaces, creating an underexamined pathway for sensitive user prompts to leak to third parties.

Why it matters

This development matters because it demonstrates that privacy vulnerabilities in AI deployments often stem from routine web application architecture rather than model-level failures. Organizations interacting with or hosting AI interfaces face severe regulatory, competitive, and reputational exposure if proprietary or confidential inquiries are harvested by third-party tracking networks. Consequently, data governance frameworks must account for client-side exfiltration vectors alongside backend API security.

Key insights

• AI chatbot platforms routinely collect intimate and sensitive user information under an assumed expectation of user privacy.
• Web interfaces of AI chat applications embed third-party analytics, advertising, and session replay scripts that function concurrently with chat functions.
• Embedded client-side scripts establish an unmonitored channel for the leakage of conversational data and user prompts to external third parties.
• An audit across 42 platforms confirmed that contextual leakage risks exist across general-purpose systems as well as sensitive domains, specifically health, legal, and educational platforms.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2610.37467