After researchers pulled 6,800 private Muse chats, Meta suspends sharing

Meta has pulled the plug on Muse's chat-sharing feature, and the reason reads like a case study in why AI features deserve a security review before they ship. Researchers gained access to 6,800 shared conversations with the AI agent, and among those chats were internal company documents. That figure, 6,800, is not a rounding error from a single test group - it is a large enough sample to suggest the exposure was systemic, not a one-off glitch. Meta responded by suspending the feature, an acknowledgment that the tool's promise of easy collaboration came with a serious privacy flaw attached to it.

The first question executives will ask is simple: how did internal company documents end up in conversations that third-party researchers could reach? The source does not say exactly what went wrong, and it does not name the researchers or the method they used. What the reporting makes clear is that the sharing feature, presumably designed to let users pass interesting or useful AI conversations to colleagues, instead became a conduit for sensitive material. For any company deploying generative AI, it is a stark reminder that the feature you add to increase productivity can also become the window that exposes your internal playbook.

For context, Meta has been embedding AI agents across its sprawling product ecosystem: Facebook, Instagram, WhatsApp, and its Ray-Ban smart glasses partnership. Muse, the assistant at the center of this incident, is part of that broader push to make Meta a daily AI destination rather than just a social media company. The company has moved fast to ship AI features, and shipping fast often means security teams are playing catch-up. This suspension shows that even Meta, with its enormous engineering and trust-and-safety resources, can stumble on something as fundamental as who gets to see an AI conversation.

The stakes go well beyond one feature on one platform. Shared chats are increasingly common in AI tools used inside companies - teams use them to hand off research, share outputs, and build on each other's prompts. But those chats can contain proprietary strategy, customer data, code snippets, and candid internal discussions. In the wrong hands, a leak like this is not just embarrassing; it is potentially a regulatory problem. Privacy regulators on both sides of the Atlantic have been scrutinizing how AI products collect, process, and expose personal data. An incident of this scale could draw attention to the broader practice of using consumer AI tools for internal corporate work.

The strategic lesson for boards and operating teams is sharp: AI features are data flows, and data flows need access controls. The feature that lets you share a useful answer with a colleague also lets you accidentally share the answer with the entire world if the permission model is wrong. This incident should push every company to inventory which internal information is being fed into AI tools, who can retrieve it, and whether those retrieval paths have been tested by people outside the security team. If a pair of researchers can find 6,800 conversations, so can a cybercriminal or a competitor.

For Meta, the calculus now is about trust. The company wants users to treat its AI assistant as a reliable partner for work and life. When a feature gets suspended after exposing internal documents, it reinforces the opposite impression: that the product is moving faster than its safety rails. Meta will likely quietly fix the feature and bring it back, but the momentum cost is real. Every day the feature stays dark is a day users find another way to do what Muse once promised, and a day competitors can point to Meta as an example of careless AI deployment.

For every other company building or buying AI tools, the play is more direct. Treat sharing features as high-risk surfaces. Audit them before launch, red-team them after launch, and assume that any conversational data synced to a cloud service is one misconfiguration away from exposure. The 6,800-chat incident is not Meta's alone to own; it is a preview of the failures that will happen across the industry unless executives start asking their AI teams the harder questions: What is in these chats? Who can see them? And what happens when a third party finds a way in?