Claroc finds 6,800 links to Meta Muse chats — Ars Technica
Researchers at Claroc said they had found 6,800 links created by users of the Meta Muse AI agent to share conversations. Some of these pages contained personal or confidential business information, Ars Technica reports.
How the links were found
Muse allows users to share searches, chats, images, and other content through web links. Typically, such addresses contain a long unique string of characters so that outsiders cannot easily guess them. However, according to Claroc, Muse links add the conversation title to the domain, making them easier to find in passive DNS databases.
The researchers provided simplified examples of addresses with topics such as a home-sale budget, a quarterly revenue review, or a draft loan application. Claroc called such pages GhostShares — they may be accessible to people who were not directly sent the link.
More current news is available on the UA.News Telegram channel Telegram.
What data the chats contained
After reviewing several hundred links, Claroc specialists found notes from a candidate preparing for a confidential call with an AI lab regarding an executive search. They also saw an internal revenue review by a marketing company and a small business owner's loan application.
Other examples included a to-do list with a bank fraud warning, an internal real estate investor report with financial metrics, and an annual employee vacation schedule. Claroc said it had checked similar platforms, consulted industry colleagues, and studied Meta Muse documentation. In the company's view, this service behavior is part of its design, but users should be aware that a link created to share a chat can be discovered through passive DNS.
A Meta spokesperson did not immediately answer Ars Technica's questions by email. Claroc co-founder and chief technology officer Asaf Aproszper said the company had also contacted Meta but received no response. The researchers advise against adding unnecessary personal information to chats intended for sharing and recommend deleting links after use.
