MENLO PARK, CALIFORNIA / RankWire.AI / – Meta has put a pause on sharing conversations through its Muse AI assistant after cybersecurity experts discovered that over 6,800 shared chat links could be retrieved via public internet records. These links, which may have revealed sensitive personal and business details, prompted the company to take immediate action. Meta informed Ars Technica on Tuesday, October 20, that the sharing feature is temporarily suspended while the company assesses the security implications. Additionally, Meta indicated it is contemplating modifications to the way shared links are made accessible.
Claroc, a cybersecurity firm, uncovered more than 6,800 shared Muse links by examining passive DNS records—databases that log domain name queries and enable web address searches. This number reflects the total discovered link addresses, not the count of sensitive chat sessions. Claroc published its findings on October 18, stating it initially reported the issue to Meta’s bug bounty program on October 8, providing detailed steps to reproduce and examples of shared links.
Researchers accessed several hundred conversations, with nearly all active shares loading without requiring a Meta account. Some of these chats contained confidential information, such as an internal revenue review for a marketing firm and a small-business loan application. Others displayed a bank fraud warning, including a transaction amount and partial account details. The researchers also found sensitive interview notes from an executive search and a vacation calendar listing dozens of employees. The remaining shared chats mostly included routine content.
Flawed link design made chats easily discoverable
The vulnerability originated from how Muse generated its sharing URLs. Each link combined the conversation’s title with a randomly created identifier within the hostname of the web address. This identifier was sufficiently complex to prevent guessing. However, when a link was accessed, it triggered a DNS lookup that passive DNS services could log and index. Consequently, third parties could locate these addresses without direct access and view chats without signing in.
Although Muse instructed search engines and AI crawlers not to index its shared pages, this did not prevent DNS record access. The security website urlscan.io stored over 75 saved scans of Muse shared links. Claroc stated it found no evidence of misuse but did not contact affected users, nor could it determine whether others had discovered the links. Meta’s official statement did not specify if users would be notified of the breach.
Sharing features disabled after review
On October 20, Claroc re-examined 40 shared links that previously opened without requiring an account. All of these links now returned errors, and the content-serving addresses were inaccessible. Nonetheless, conversation titles remain stored in passive DNS logs. Meta launched Muse in the United States on September 8, targeting users aged 18 and older. The assistant is capable of connecting to users’ email, calendar, payment, and health applications. In April, Meta introduced its Muse Spark model developed at its superintelligence research center.
This incident underscores rising concerns around data privacy in AI-driven tools, especially regarding the sharing of confidential information. Similar vulnerabilities were seen in 2025 when shared ChatGPT chats surfaced in Google search results. Meta also faces legal challenges related to user safety, including being ordered to pay $567 million following a New Mexico court ruling in August. The company has not announced a timeline for resuming Muse chat sharing.