Meta has temporarily suspended a sharing feature in its Muse AI assistant after security researchers discovered that thousands of shared conversations could be found online. The exposed material reportedly included confidential business documents and personal financial information.
According to Ars Technica, cybersecurity firm Claroc uncovered the issue after identifying more than 6,800 web addresses associated with conversations shared through Muse. Researchers were able to access hundreds of these pages without signing into a Meta account, finding information that users would not typically want publicly available.
The exposed content included an internal revenue review belonging to a marketing company, a small business owner's loan application, a financial report prepared by a real estate investor and a workplace calendar detailing employees' planned vacations. Researchers also discovered notes prepared for a confidential conversation with a senior figure at an AI laboratory.
Meta has since paused the sharing feature while investigating the findings and considering changes to make the links more difficult to discover.
Muse spokesperson Tracy Clayton told The Verge that users could share specific items generated through the assistant, creating web pages accessible to anyone possessing the corresponding link. The company said it had suspended the functionality while reviewing the report and evaluating potential changes to how these links could be discovered.
What happened and how?
Muse enables users to share conversations, images and other content through web links. This allows them to send AI-generated material to colleagues and other recipients. However, Claroc identified a weakness in the way these links were generated.
The web addresses contained a brief description of the conversation followed by a randomly generated identification code. As a result, the addresses could reveal information about the contents of a shared page even before it was opened.
Researchers used passive DNS records, which preserve information about domain names observed by internet infrastructure, to locate these addresses. When someone accessed a shared page, its URL could appear in searchable records. Researchers could then use the available information to reconstruct links and open the associated conversations.
Claroc said it discovered more than 6,800 addresses through a single search query. However, the researchers examined only a few hundred links instead of opening every address they identified.
The firm submitted its findings to Meta's bug bounty programme on October 8. Claroc co-founder and chief technology officer Asaf Aprozper also reportedly tried to contact Meta employees through LinkedIn, according to Ars Technica.
What is Muse AI?
Launched on September 8, Muse is a general-purpose AI assistant designed to perform tasks such as completing forms, booking appointments, making purchases and creating documents. Carrying out these activities can require access to users' accounts and personal information.
The discovery has raised questions about the security of shared AI conversations, particularly when people use these assistants to discuss financial matters, workplace details or confidential business strategies.
Claroc said it had found no evidence that the exposed conversations had been misused. The firm also clarified that its investigation focused on one method of discovering shared pages and did not establish the full scale of the exposure.
The incident comes after previous cases in which shared conversations from AI assistants, including ChatGPT and Grok, appeared in Google Search results.
