Meta Muse exposed a privacy vulnerability: a single search found over 6,800 chat links, and corporate secrets were also uncovered.

Beating AI News Flash: U.S. cybersecurity firm Claroc has discovered that chat share links from Meta's personal AI Agent Muse can be bulk-retrieved by strangers. Researchers queried a DNS history database just once and found more than 6,800 share links. After spot-checking hundreds of them, they found dozens of sensitive documents, including internal corporate revenue reports, small business loan applications, and preparation materials for a candidate's confidential recruitment communications with a frontier AI lab — even internal contacts and meeting times were exposed.

When users ask Muse to share a chat, the system generates a webpage that can be viewed without logging in, and writes the chat title together with a random ID into the domain name. Once someone opens the page, the full domain name may be recorded by third-party DNS services. Claroc simply searched these historical records and obtained share addresses that would otherwise be difficult to guess. Although Muse asks search engines not to crawl these pages, it cannot prevent domain names from being recorded by other services.

Claroc reported the vulnerability to Meta as early as October 8 and published its research on October 18. It was not until October 20, after U.S. tech media outlet Ars Technica intervened with inquiries, that Meta confirmed it had suspended the chat sharing feature. That same day, Claroc rechecked 40 links that had previously been accessible without logging in and found that all had become invalid, but the chat titles remained in DNS history records.

Meta CEO Zuckerberg said last month that Muse is a product "designed from the ground up for privacy and security."