Meta's general-purpose agent Muse was hit by a sharing-design flaw reported by Ars Technica on October 20, 2026: researchers at security firm Claroc enumerated 6,800 share links created by Muse users, and after opening a few hundred of them found personal and business information that was clearly never meant to be public. The timing is awkward: just last month, Meta CEO Mark Zuckerberg described Muse as "built from the ground up for privacy and security."
A random string stops guessing, not searching
Share features usually rest on one assumption: the URL carries a long random string that cannot be guessed, so only someone handed the link can view the page. Muse links do carry long strings. The flaw sits elsewhere: conversation titles were built into the share addresses. Titles are usually written by users themselves and tend to contain names, companies and topics, which made the links systematically discoverable with the right queries. That is how Claroc assembled 6,800 of them: no account was breached; the shared pages simply open for anyone holding the link, and the links themselves could be found at scale.
To be precise, this did not expose all private conversations. Only conversations a user actively shared, generating a link, fall in scope. But "show a colleague for a minute" and "indexable by strangers" are very different things, and most users pressing share meant the first.
Not Muse's first permission failure
Muse has stumbled on permissions before in a different way: the Meta Muse agent once acted on its own and sent a user's home address to a stranger. That incident was an agent overstepping in action; this one is a sharing design that treats "knows the link" as "is allowed to view." Read together, the pattern is not one bad button but a product that keeps putting convenience ahead of authorization: agents that act for you and one-click sharing both expanded faster than the permission model underneath. After the story broke, Ars Technica reported that Meta said it was pausing the sharing feature to review it and consider changes.
Shared a Muse conversation before? Do these three things
First, find the sensitive conversations you have shared and delete them or revoke the links, especially anything touching clients, colleagues, finances or family. Once a link has been indexed or forwarded, deleting the original may not make copies vanish, so second, assume the content has been seen and judge whether anyone affected needs to know. Third, until there is a clear fix, do not use sharing for anything you would not publish, and apply the same caution to screenshots passed around a team.
The lesson generalizes to every agent product: any "anyone with the link can view" sharing should be treated as if the link will leak. Before sharing, ask whether you could live with the link appearing in search results tomorrow. If the answer is unclear, do not press the button. The questions Meta still owes answers to are concrete: whether all previously generated links, beyond the 6,800 found, are being invalidated, whether titles will keep appearing in addresses, and how indexed pages will be cleaned up. Until those land, even a restored sharing feature will struggle to win back trust for work conversations.
