Meta has temporarily suspended the conversation-sharing feature in its Muse AI assistant after cybersecurity researchers discovered more than 6,800 shared conversation links that could be found through public internet records. The discovery has raised concerns about how artificial intelligence platforms create, share and protect conversations that may contain sensitive personal or business information.
The issue was reported by Claroc, a cybersecurity firm whose researchers said they located the links through publicly available Domain Name System records. After examining hundreds of pages, the researchers found examples containing financial details, internal company documents, workplace schedules and confidential meeting notes.
Meta confirmed that it had paused the sharing functionality while reviewing the findings and considering changes to how shared links could be discovered. The incident highlights the privacy risks that can arise when AI services allow users to share conversations through web links without sufficiently limiting the discoverability of those links.
However, the findings do not mean that every Muse conversation was exposed. The investigation concerned conversations that users had chosen to share through links. Researchers also said they had no evidence that the discovered conversations had been misused.
How Were More Than 6,800 Links Discovered?
According to Claroc, the researchers identified more than 6,800 Muse conversation-sharing links by searching public DNS records. DNS, or the Domain Name System, helps internet-connected devices locate websites and online services by translating domain names into the information needed to connect to them.
Some DNS monitoring services retain historical records of domain names and related information. Researchers can examine these records to investigate online infrastructure and identify patterns that might otherwise be difficult to find.
Claroc reported that Muse's shared conversation URLs used a structure that incorporated the title of a conversation alongside a random identifier. Because the conversation title appeared in the web address, the researchers could identify the subject of a page from the hostname and reconstruct the corresponding link.
The researchers said this design made the shared pages easier to discover through passive DNS searches. They found that a single search could reveal thousands of links associated with the service.
The concern was not simply that someone possessing a shared link could open a conversation. The problem was that the links could be discovered without being directly provided by the person who created them.
What Information Was Found in the Shared Conversations?
Claroc said it examined hundreds of the discovered pages and found several examples containing information that users would not normally want to make widely accessible.
The reported examples included a marketing company's internal revenue review, a small-business owner's loan application and a workplace calendar listing employees' planned leave. Researchers also found notes prepared for a confidential meeting related to an executive recruitment process.
Another example reportedly contained a task list referring to a bank fraud alert, including a partial account number and transaction details. A real-estate investment report also contained property addresses and financial figures.
These examples demonstrate how AI conversations can contain information that extends beyond ordinary questions and answers. Users may upload documents, describe financial circumstances, discuss workplace matters or ask an AI assistant to prepare reports using confidential information.
If a shared page becomes discoverable beyond its intended audience, the information it contains could create privacy or business risks. Depending on the details exposed, it could potentially be used for targeted scams, impersonation or unauthorised disclosure of commercial information.
However, the examples do not establish that every discovered page contained sensitive information. Researchers reported that many pages they examined were harmless, including ordinary business information, listings and event calendars.
Why Did Muse's Link Design Raise Concerns?
Muse allowed users to create web links to share selected AI conversations and related content with other people. This type of functionality is common across online services because it makes information easier to send to colleagues, friends or business partners.
The security concern arose from how the links were structured. According to Claroc, conversation titles were included in the hostnames used for shared pages. Public DNS records could therefore reveal identifying information about the shared content, making the links easier to locate.
A random identifier can help prevent someone from guessing a URL, but it is not a complete privacy safeguard if other elements of the address make the link discoverable through public records or third-party indexing services.
The incident illustrates the difference between a link that is not prominently advertised and content that is genuinely restricted to authorised users. If a page can be opened by anyone who obtains its URL, its privacy depends partly on keeping that URL from being discovered or shared further.
For sensitive information, access controls that require authentication or explicitly restrict the intended audience can provide stronger safeguards than relying only on an obscure web address.
Meta Pauses the Sharing Feature
Meta spokesperson Tracy Clayton told The Verge that Muse's sharing functionality created pages that could be accessed by people who had the relevant link. The company said it had paused the feature while reviewing the research and considering changes to how links could be discovered.
The pause is intended to give Meta time to assess the reported issue and consider adjustments to the design of its sharing system. The company has not publicly established through the cited reporting that every discovered conversation was accessed by an unauthorised person before the suspension.
Claroc said it reported the issue to Meta's bug bounty programme on October 8, 2026. The firm also said it was withholding the discovered hostnames, titles and links rather than publishing information that could make the conversations easier to access.
The investigation shows why responsible vulnerability reporting is important. Researchers can demonstrate a weakness and alert the platform operator without distributing the potentially sensitive material more widely.
Does This Mean All Muse AI Chats Were Leaked?
No. It is important to distinguish between shared conversations and all conversations created with Muse.
The reported discovery involved links created through Muse's sharing functionality. It does not establish that all private chats were publicly available or that every Muse user was affected.
The researchers identified more than 6,800 links through one discovery method and examined hundreds of pages. Claroc said its findings did not measure the total number of shared conversations or establish how long every link had been accessible.
The firm also reported no evidence that the conversations it discovered had been misused. This does not prove that misuse was impossible, but it means the reported investigation did not establish that someone had exploited the exposed information for fraud or other malicious activity.
Users should therefore avoid interpreting the incident as confirmation that every conversation or account was compromised. The key concern is that some conversations deliberately shared through links were easier to discover than users may have expected.
What Should Muse Users Do?
Users who have shared Muse conversations should review the content they made available and consider whether any links contain sensitive information.
Where the service provides a way to revoke or delete shared links, users should consider using it for conversations that no longer need to be accessible. Because Meta has paused the sharing feature while reviewing the issue, the available controls may change as the company updates the service.
Users should also avoid including unnecessary personal, financial or confidential business information in conversations that they intend to share. Before sending a link to someone else, they should check what content it contains and understand whether the recipient can forward it or access it without signing in.
If a shared conversation contains banking information, passwords, authentication codes or other credentials, users should take appropriate steps to protect those accounts. For example, they may need to change exposed passwords or contact the relevant financial institution if there is a credible risk that account information has been compromised.
Businesses should establish clear rules about which AI tools employees may use for confidential documents, client information and internal financial reports. Organisations should also check whether their AI platforms offer appropriate access restrictions, administrative controls and data-handling policies.
A Wider Challenge for AI Privacy
The Muse incident adds to a broader debate about privacy in AI assistants. People increasingly use these tools to prepare reports, summarise documents, plan business activities and organise personal information. As a result, AI conversations can contain details that would traditionally be stored in private emails, internal documents or business applications.
Sharing features can be useful, but they also introduce risks if users do not understand how the resulting links work. A link designed for convenience can become a privacy concern when it is discoverable through public records or when anyone possessing it can access the underlying content.
Technology companies need to consider not only whether a link is difficult to guess, but also whether the link's structure reveals information, whether access can be restricted and whether users receive clear explanations before sharing potentially sensitive material.
For users, the incident is a reminder to treat AI conversations with the same care as other digital documents. Sensitive information should be shared only when necessary, and users should understand the visibility and access settings of the service they are using.
What Happens Next?
Meta has paused Muse's conversation-sharing functionality while reviewing the findings from Claroc and considering changes to the way shared links are discovered. Further details about any redesign or the restoration of the feature will depend on the company's subsequent announcements.
For now, the reported incident highlights the importance of secure link design and clear privacy controls in AI products. The discovery of more than 6,800 links demonstrates that content intended for sharing with a limited audience can sometimes become easier to find than expected.
The findings do not establish a leak of all Muse conversations or confirm that the discovered information was misused. Nevertheless, they underline why AI companies must design sharing tools with privacy in mind and why users should carefully review sensitive information before making a conversation accessible through a link.
