MosaicLeaks is an experiment that tests the confidentiality of AI-powered research agents. These agents are designed to assist researchers by gathering and summarizing relevant papers, but they may also have access to sensitive information. The MosaicLeaks project aims to see if these agents can keep secrets when asked not to disclose them. --- Why it matters: This matters because AI-powered research assistants may soon be ubiquitous in academia and industry, handling confidential data and making decisions based on it. If these agents cannot maintain confidentiality, it could have significant implications for research integrity and security. Source: https://huggingface.co/blog/ServiceNow/mosaicleaks