MENLO PARK, CALIFORNIA / RankWire.AI / – Meta has temporarily halted the sharing of conversations within its Muse AI assistant following cybersecurity researchers’ discovery that thousands of shared chat links could be retrieved via public internet records. The exposed links potentially revealed sensitive personal and corporate data. On Tuesday, October 20, Meta informed Ars Technica that it had paused this feature while it investigates the findings. The company also mentioned it is contemplating modifications to how shared links are discovered.
During recent analysis, security firm Claroc identified over 6,800 shared Muse links through passive DNS records. These records, which log domain-name lookups, make web addresses searchable. It’s important to note that this figure represents the number of discovered link addresses, not the total number of sensitive conversations. Claroc released its research on October 18, stating that it initially reported the issue to Meta’s bug bounty program on October 8, providing reproduction steps and sample shares.
By opening several hundred conversations, researchers found that nearly all live shares could be accessed without a Meta account. Some of these contained sensitive material, such as a marketing firm’s internal revenue review and a small-business loan application. Others displayed a bank fraud alert with a transaction amount and partial account details. Researchers also uncovered confidential interview notes from an executive search and a workplace vacation calendar listing numerous employees. Many other shared chats consisted of routine content.
Vulnerable link structure made chats discoverable
The vulnerability originated from how Muse generated its share links. Each URL combined the conversation’s title with a randomly assigned identifier within the hostname. This identifier was sufficiently long to prevent guessing. Nevertheless, when a link was opened, it triggered a DNS lookup, which passive DNS services could record and index. Consequently, others could locate these addresses without direct access and view the chats without signing in.
Although Muse instructed search engines and AI crawlers not to index its shared pages, this did not prevent access via DNS records. Additionally, the security site urlscan.io stored more than 75 saved scans of Muse share pages. Claroc stated it found no evidence of misuse, and did not contact the affected users, nor could it determine if others had found the links. Meta’s official statement did not clarify whether users affected by this issue would be notified.
Sharing functionality halted following review
On October 20, Claroc re-examined 40 shared links that had previously loaded without a Meta account. None of these accessed the content, and the servers returned errors. Despite this, conversation titles still appeared in passive DNS history. Meta rolled out Muse in the United States on September 8, targeting users aged 18 and older. The assistant integrates with users’ email, calendar, payment, and health apps. In April, Meta introduced its Muse Spark model from its superintelligence laboratory.
This episode underscores the increasing worries over data privacy within AI tools and the danger of sharing conversations containing confidential details. Similar issues surfaced as early as 2025, when shared ChatGPT chats appeared in Google search results. Moreover, Meta faces legal challenges related to user safety, including a ruling that mandated the company to pay $567 million following a New Mexico court decision in August. Meta has yet to announce when Muse’s sharing feature will be restored.