It’s an AI agent of chaos.
Meta’s allegedly ironclad AI assistant Muse sparked privateness issues after it was caught exposing customers’ delicate and business info.
Ahead of the Muse’s release last month, Meta CEO Mark Zuckerberg had burdened the “privacy and security” of the cyber secretary, which executes duties on the customers’ behalf, from sending emails to reserving journey, Futurism reported.
However, an investigation by security firm Claroc discovered that the supposedly watertight tech was anything but. The company’s investigators reportedly found a glitch that allowed them to access pages from over 6,800 hyperlinks Muse had created so customers could share pages with each other, Ars Technica reported.
Mark Zuckerberg, CEO of the Facebook-owned Meta Group, unveils a small machine called Muse Charm designed to present access to the company’s proprietary AI software program, Muse, at the Meta Connect developer convention in Menlo Park, United States, Wednesday, Sept. 23, 2026. ZUMAPRESS.com
These included loan purposes with the requested quantity, bank transaction particulars, a real estate investor’s inside efficiency report, and a company income doc meant for the CEO.
There was even a company’s full-year trip calendar itemizing the PTO dates of dozens of workers.
Worst of all, these secrets and techniques were unveiled with the push of a button. “If you had asked us beforehand, we would have guessed that gathering these would take real effort,” the researchers wrote. “It took one query.”
To access this intimate information, the investigators reportedly exploited Muse’s share perform, which permits customers to share searches, conversations, pictures, and other content material among one another.
A meme that includes Jolly’s, Muse’s mascot. Alexandr Wang/X
This content material is accessed via AI-generated web hyperlinks that usually comprises a long, cryptic string of characters that theoretically make it arduous to hack.
However, the Claroc staff discovered Muse’s URLs apparent as they all used the area title cf.metaaiusercontent.com as effectively as the AI-generated dialog title alongside the textual content sequence.
“If you had asked us beforehand, we would have guessed that gathering these would take real effort,” the researchers wrote. “It took one query.”
Examples included home-sale-budget-9f21c8a3.cf.metaaiusercontent.com and loan-application-draft-a4c9e2.cf.metaaiusercontent.com.
“We call these GhostShares: pages that open for people who were never handed the link, and that those people can find on their own,” wrote Claroc.
In mild of the oversight, Muse has shut down the characteristic. “We’ve paused this functionality while we review this report and consider changes to how these links can be discovered,” company reps informed Ars in a assertion.
However, this wouldn’t be the first time someone has allegedly found a privacy-violating glitch in the Meta-verse.
Last week, Meta glasses wearers claimed that the cutting-edge tech specs were listening to them even when set to sleep mode.
“Meta AI does not respond to me, and I never intentionally activated it,” lamented one Redditor. “I have observed the white notification mild activating while I am talking usually.
Source hyperlink
