Meta's Muse AI exposes sensitive user information, suspends sharing feature

Meta has suspended a sharing feature in its Muse AI assistant after security researchers found that thousands of shared conversations could be discovered online, including documents containing confidential business information and personal financial details. Ars Technica reported that the issue was discovered by cybersecurity firm Claroc, which found more than 6,800 web addresses linked to conversations shared through Muse. Researchers were able to open hundreds of these pages without logging into a Meta account and found several that contained information that users would not normally want publicly accessible.The exposed material included a marketing company's internal revenue review, a small business owner's loan application, a real estate investor's financial report and a workplace calendar listing employees' planned vacations. Researchers also found notes prepared for a confidential call with a senior figure at an AI lab. Meta has now paused the feature while it reviews the findings and considers changes to prevent these links from being discovered so easily."People have the ability to share specific Muse artifacts, which creates a page that can be accessed by those with the link," Muse spokesperson Tracy Clayton told The Verge. "However, we've paused this functionality while we review this report and consider changes to how these links can be discovered."What happened and howMuse allows users to share conversations, images and other content through web links. Anyone who receives a link can open the shared page, making it easier to send AI-generated material to colleagues or other people. However, Claroc found a problem with how Muse created these links. The web addresses included a short description of the conversation, followed by a random identification code. This meant the addresses could reveal what a shared page contained even before someone opened it. Researchers used passive DNS records, which store information about domain names observed by internet infrastructure, to find these addresses. Once a person opened a shared page, its web address could appear in searchable records. Researchers could then use the information to reconstruct the link and access the conversation.According to Claroc, finding more than 6,800 such addresses required just one search query. The researchers said they examined only a few hundred links, rather than opening every address they found. The company reported its findings to Meta's bug bounty programme on October 8. Claroc co-founder and chief technology officer Asaf Aprozper also attempted to contact people at Meta through LinkedIn, according to Ars Technica.What is Muse AI?Muse launched on September 8, as a general-purpose AI assistant designed to carry out tasks such as filling out forms, booking appointments, making purchases and creating documents. These capabilities can require access to a user's accounts and personal information.The latest discovery raises concerns about how shared AI conversations are protected, particularly when users discuss financial matters, workplace information or confidential business plans with an assistant.Claroc said it found no evidence that anyone had misused the exposed conversations. It also cautioned that its findings covered one method of discovering shared pages and did not establish the full extent of the exposure.The incident follows earlier cases in which shared conversations from AI assistants, including ChatGPT and Grok, appeared in Google Search results.- Ends