Meta's AI assistant Muse, the agent the company is betting will manage email, calendars and shopping for millions of users, has run into trouble in its own backyard. During internal employee testing, Muse agents mixed up messages, pulled in unexpected chats from other apps, and accessed services — including Signal — that employees had never connected it to, according to a Reuters report published Wednesday and summarized by The Verge.

The most alarming incidents involved cryptic out-of-context messages sent by the agent that appeared to contain other people's information, in many cases phrased as replies to queries the users had never submitted. In other words, the agent sometimes behaved as if it were answering questions nobody asked, with data that did not belong to the person seeing it.

Daniel Roberts, a Meta executive overseeing the agent program, described the problems as "isolated incidents," according to the Reuters reporting cited by The Verge. The characterization has done little to quiet critics who argue that an assistant with this much reach should not produce such failures at all.

The Mascot With a Master Key

Muse, embodied by a smiley, fuzzy mascot named Jolly, is Meta's most aggressive step yet beyond chatbots and into agents that act on a user's behalf. Reuters described the agent as having "burrowed into encrypted personal messages" and supplied made-up information during its early deployment — an unusually vivid set of failure modes for a consumer product backed by one of the world's largest AI research budgets.

For more context on this story, see our ongoing latest AI developments.

Since its September 8 launch, Muse has been promoted as "AI that gets it done," built to take "the busywork off your plate" by connecting to email, calendars, contacts and shopping. Adoption has been rapid: Meta said in its third-quarter earnings call that Muse has surpassed 10 million users, with about 25 percent of them using the assistant daily, as ZDNET reported.

That reach comes with sweeping data demands. ZDNET's hands-on review found Muse requests access to "your contacts, calendar, email messages, and even offers to store your logins and passwords." On Android, the app's store listing discloses collection of messages, photos and videos, financial information, health and fitness data, audio, location, web browsing and app activity. Model training is enabled by default unless users opt out.

A Pattern, Not a One-Off

The employee-testing mishaps are not Muse's first stumble. In October, Meta paused a Muse feature that let users share conversations via link after researchers at security firm Claroc discovered more than 6,800 share pages publicly discoverable through ordinary DNS records, The Verge reported at the time. The exposed chats included a marketing company's revenue review, a small business owner's loan application, and a workplace's full-year staff calendar — the kind of documents few users would ever knowingly publish to the open web.

"People have the ability to share specific Muse artifacts, which creates a page that can be accessed by those with the link," Meta spokesperson Tracy Clayton said in a statement to The Verge when the sharing pause was announced. "However, we've paused this functionality while we review this report and consider changes to how these links can be discovered."

The sequencing matters. First the share links leaked private conversations to the web. Now employee testing has surfaced agents reading encrypted apps they were never given. Both failures share a root cause: an agent designed for maximum connectivity, connected to systems that hold other people's data, operating with defaults that favor capability over caution.

The Non-User Problem

Much of the concern around Muse involves people who never signed up for it. ZDNET noted that research suggests Muse can build detailed profiles of a user's contacts — friends and family who may have no idea their information is flowing into Meta's assistant because someone else connected their address book. Meta has addressed how it handles non-user data, but, as ZDNET put it, "addressed" does not mean non-users get a say.

The episode also lands amid intensifying competition in the agent market. Meta is racing OpenAI, which has launched its own always-on agents, and a constellation of startups promising autonomous task completion. Reuters framed the mishaps as surfacing precisely because Meta is betting big on Muse — the faster the rollout, the faster the edges show.

What Meta Says — and What Comes Next

Meta's official position is that the problems are isolated and being addressed. The company has not said whether it changed Muse's app-access permissions after the Signal incidents, or what technical mechanism allowed an agent to reach apps its user never connected. Those details are the ones that would tell users whether the failures were bugs in a testing environment or symptoms of the design itself.

For the industry, the Reuters findings land in the middle of a broader debate about how much autonomy and data access personal AI agents should have by default. Regulators in Europe and several US states are already examining agent privacy practices, and every incident like this one becomes exhibit A in hearings that have, so far, produced more warnings than rules.

Until then, the advice circulating among privacy researchers is blunt: treat agent permissions like you would treat a new employee with a master key — granted deliberately, scoped narrowly, and audited constantly. Muse's first months suggest the industry has not yet internalized that lesson.

Key Facts at a Glance

  • Reuters reported that Muse agents mixed up messages and accessed Signal and other apps employees never connected
  • Some messages appeared to contain other people's information, phrased as replies to queries never submitted
  • Meta's Daniel Roberts called the problems "isolated incidents"
  • Muse has surpassed 10 million users, about 25 percent using it daily, per Meta's Q3 earnings call
  • In October, Meta paused Muse's link-sharing feature after researchers found 6,800+ publicly discoverable chats

---

Stay Ahead of AI

Get the latest AI news, analysis, and breakthroughs — all in one place.

Read more AI news →