CLOAK: A Privacy-Preserving Assistant for Online Social Media Users (Video Preview)

CLOAK: A Privacy-Preserving Assistant for Online Social Media Users Sandeep Kalari, Mohan Sunkara, Vikas Ashok, Ravi Mukkamala UIST 2026 Video Preview for a UIST 2026 paper Many people share personal details on platforms like Reddit under the assumption that a pseudonymous account keeps their identity safe. They do not realize that a profession mentioned in one subreddit, a hobby in another, and a neighborhood in a third can be pieced together to narrow down exactly who they are. Large language models have made this kind of cross-post profiling easy to do at scale. Yet existing tools offer little help: PII scanners catch names and emails but miss contextual cues, and even LLM-based privacy systems analyze only the current post in isolation, missing the cross-post combinations that actually compromise anonymity. We present CLOAK, a system that examines a user’s full posting history alongside their draft comment and shows them the identification risk before they post. A pipeline of four LLM agents extracts a structured attribute profile from prior posts, identifies cross-post interactions that narrow the anonymity set, ranks risky phrases, and suggests minimal edits that preserve the user’s intent. CLOAK is designed as an awareness tool for informed self-disclosure: rather than blocking or redacting content, it helps users understand the trade-offs and decide for themselves. In a technical evaluation on 368 Reddit profiles, CLOAK detected cross-post tipping points in 77% of cases while the status quo baseline detected none. In a user study (N =20), CLOAK was unanimously selected as the most helpful tool (trust: 6.5/7 vs. 1.5/7 for PII scanning), and every participant reported changed posting behavior. Web:: https://programs.sigchi.org/uist/2026/program/content/262913