The proliferation of AI conversation platforms has introduced unprecedented privacy risks through user-shared conversations. This paper presents a comprehensive analysis of privacy vulnerabilities in shared conversations across three major LLM platforms: ChatGPT, Microsoft Copilot, and Google Gemini. We collected and analyzed 100 342 conversations using an automated LLM-based privacy detection pipeline enhanced with a defined risk scoring system and the LINDDUN threat modeling framework. Our analysis identifies 8 131 conversations (8%) to incur privacy risks deriving from the disclosure of private and sensitive data including user identifiers (49%) and user location data (40%), yet in some cases also financial (4%), health (3%) and authentication data such as access tokens (3%). Through systematic analysis of conversation length and temporal disclosure patterns, we demonstrate that extended conversations exhibit higher privacy risk rates compared to brief interactions. Notably, 60% of private data disclosures in longer con- versation occur in the final quartile of these conversations, which may indicate that users progressively lose privacy awareness as interactions deepen. Our findings have immediate implications for platform designers and policymakers, highlighting the need for proactive interventions including real-time privacy warnings, pre- share scanning, and clearer education about the permanence and discoverability of shared conversation links.

A web visit typically consists of the browser rendering a dynamically generated response that is specifically tailored to the user. This generation of responses based on the currently authenticated user, whose authentication credentials are automatically included via cookies in all (including cross-site) requests, have led to a multitude of issues. Through cross-site leaks (XS-Leaks), an adversary can try to circumvent the same-origin policy and extract information about responses, which in turn can reveal potentially sensitive information about the user. As research on this class of vulnerabilities only recently gained traction, and the attacks affect many different components of the web platform, the intrinsic characteristics and underlying causes remain largely unexplored.

In recent years, e-books have proven to be a very appealing alternative to physical books; nowadays, almost every written book is published in an electronic format next to its physical copy. In an attempt to promote consensus and to offer an alternative to emerging proprietary e-book formats, the Open eBook format was introduced, now known as the EPUB format. Building on existing web functionalities, this open format relies primarily on XHTML and CSS to construct e-books. As such, browser engines are often employed to render the contents of EPUBs. However, this implies that reading systems may face similar vulnerabilities as web browsers.In this paper, we report on a semi-automated evaluation of the security and privacy aspects of EPUB reading systems. This evaluation, which was performed on 97 EPUB reading systems covering seven platforms and five physical reading devices, revealed that almost none of the JavaScript-supporting reading systems sufficiently adhere to the EPUB specification’s security recommendations. Furthermore, our results indicate that 16 reading systems even allow an EPUB to leak information about the user’s file system, and in eight cases extract file contents. In addition to the semi-automated evaluation, we demonstrate that an attacker can launch even more potent attacks that may lead to a full compromise of a user’s system, by exploiting aspects specific to the implementation of reading systems used by millions of users. Finally, we investigate the root cause of the identified security and privacy issues, uncovering several flaws in both the implementation of EPUB reading system, as well as shortcomings of the EPUB specification.