Security researchers from the company "Claroc" announced the discovery of a security vulnerability in the new AI assistant "Muse" from "Meta," which allows access to highly sensitive information belonging to individuals and companies. The problem lies in the way web links for sharing conversations are generated, as they were found to be easily discoverable via public search tools, despite the company's previous assurances that the tool was designed with maximum privacy in mind.
How is data leaked?
The sharing feature in Muse relies on generating a unique URL that allows anyone who possesses it to view the content of the conversation, images, or search results. Companies typically use long, complex character strings in these links to prevent them from being guessed. However, researchers found that a central design decision in the structure of these links made them susceptible to mass discovery.
The researchers used a technique known as "Passive DNS search" via available commercial services, which enabled them to identify 6,800 links shared by Muse users. Once a sample of these links was opened, data appeared that was not intended to be publicly available under any circumstances.
Examples of exposed data
The report revealed shocking content contained in the leaked links, most notably:
- Notes from a candidate for a leadership position preparing for a secret call with a leading AI lab, which included internal contact names and meeting times.
- An internal revenue review for a national marketing firm, specifically addressed to the CEO.
- A loan application for a small business owner, detailing the requested amount and business specifics.
- A to-do list including a bank alert regarding a fraud attempt, with partial account numbers and transaction details.
- An internal performance report for a real estate investor, including specific property addresses and precise income and mortgage figures.
- An annual employee vacation calendar for an employer, listing dozens of employees with their absence dates.
Contradiction with privacy promises
This discovery puts "Meta" in an embarrassing position, especially since CEO Mark Zuckerberg stated last month that Muse was "built from the ground up with a focus on privacy and security." The researchers explained that the process of collecting these links did not require significant effort as expected; rather, it was done through a single query thanks to the use of the subdomain (cf.metaaiusercontent.com) for all shared links, which facilitated the process of tracking and archiving them in public DNS databases as soon as they were clicked for the first time.
What changes practically for users?
This incident demonstrates that the "share via link" feature in AI tools carries inherent security risks if domain infrastructure is not strictly secured. For users, this means that any conversation converted into a sharing link should be considered "public" technically, even if the link is not publicly posted on social media platforms. AI tool users are advised to avoid sharing conversations that contain financial data, trade secrets, or specific personal information, and instead rely on manually copying text and securing it via encrypted communication channels to ensure it does not leak into search engines or security databases.