Concerns Arise Over Meta’s AI Assistant Muse
It appears that an AI agent meant to streamline tasks has turned into a source of chaos.
Meta’s AI assistant, Muse, recently raised alarms regarding user privacy after it reportedly revealed sensitive personal and business information.
Prior to Muse’s launch last month, CEO Mark Zuckerberg emphasized the importance of “privacy and security.” The AI is designed to handle various tasks for users, from sending emails to making travel arrangements.
However, an investigation by the security firm Claroc found significant flaws in Muse’s design. The team’s research revealed a glitch that granted them access to pages from over 6,800 links that Muse had created for users to share with one another.
Included in this trove of information were loan applications detailing requested amounts, bank transaction records, a report on real estate investment performance, and even a document outlining corporate revenue meant for the CEO.
Perhaps the most alarming revelation was a company vacation calendar that disclosed the paid time off (PTO) dates for numerous employees.
This sensitive data was extracted with alarming ease. Claroc’s researchers stated, “If you had asked us beforehand, we would have guessed that gathering these would take real effort. It took one query.”
To obtain this personal information, investigators exploited Muse’s sharing function, which enables the exchange of searches, conversations, images, and other content among users.
This content is accessible via AI-generated web links, which are designed to be cryptic. But the Claroc team found these links were not as secure as intended, all sporting the domain cf.metaaiusercontent.com along with recognizable conversation titles and sequences.
Examples of these links included home-sale-budget-9f21c8a3.cf.metaaiusercontent.com and loan-application-draft-a4c9e2.cf.metaaiusercontent.com.
Claroc dubbed these issues “GhostShares”—essentially pages that individuals could access without ever having been given the link. They noted, “These pages open for people who were never handed the link and that those people can find on their own.”
Following this discovery, Meta has disabled the feature. Company representatives stated, “We opened the Muse bug bounty program so researchers could raise different scenarios and help us continue improving the product.” While users can share specific artifacts through Muse, creating links accessible to others, this functionality has been paused for review amidst these concerns.
This isn’t the first incident raising privacy concerns in Meta’s technology. Just last week, users of Meta glasses reported that the devices seemed to be listening to them even in sleep mode.
One Reddit user expressed frustration, stating, “Meta AI does not respond to me, and I never intentionally activated it. I have noticed the white notification light activating while I am speaking normally.”
