Preliminary research
Privacy risks of agentic oversharing on the Web (SPILLAGE)
AI-collected research leads through 3 November 2026, including a November–October–September missed-article review and a targeted multilingual search alongside earlier community-source sweeps. Unranked, incomplete, not community-vetted, and subject to change.
SPILLAGE measures what LLM web agents disclose while shopping on Amazon and eBay, along two axes: explicit versus implicit disclosure, and content versus behaviour. 180 tasks over Browser-Use and AutoGen with GPT-4o, o3 and o4-mini, 1,080 runs. Task-irrelevant facts leak into search strings and into clicks and form choices, with behavioural oversharing dominating; prompt-level instructions to be private do not stop it, and stripping irrelevant input raised task success by up to 17.9%.
Record
- Researcher
- Ali Shahin Shamsabadi
- Published by
- Brave
- Date
In the archive
Tags
This page is the archive's own catalogue record. The research is the work of Ali Shahin Shamsabadi, first published at the original source. Preserved copies are kept so the citation survives its host.