Meta suspends Muse sharing after sensitive documents turn up in shared chats

Meta has temporarily halted the Muse AI assistant's capability to share conversations after researchers discovered sensitive documents and personal financial information exposed through public DNS records. The issue stems from Muse's method of generating shared conversation links, which included a descriptive title and a random identifier that could be found in searchable DNS records.

Researchers at Claroc identified over 6,800 of these shared hostnames, which could be reconstructed without needing to log in. The exposure occurred because Muse's robots.txt file instructed search and AI crawlers to avoid the share paths, but did not prevent DNS services from recording the hostnames. While some of the accessed pages were innocuous, others contained sensitive information such as a marketing company's revenue review, a small business owner's loan application, and a workplace calendar with employees' vacation schedules.

These pages did not require any special authentication to view. Claroc's investigation was limited to examining a few hundred pages and did not confirm the sharing intentions of the people or companies involved. However, no evidence of misuse was found, and it was unclear if others had discovered the links. Meta reported the issue to its bug bounty program on October 8, and after suspending the feature on October 20, none of the previously accessible shares remained so, but the descriptive information in the DNS records remained.