wasl.tech

Meta’s Muse AI Agent Is Mixing Up Conversations — and Showing Users Other People’s Private Data

It started with a fishing forecast. A Meta employee asked Muse, the company’s new AI agent, about the weather and the tides for a weekend trip. The answer came back as expected — and then Muse added something extra, unprompted.

“BTW, two bank updates, your Capital One balance is $2722.58, your Charles Schwab balance is $5899.57.”

The employee had never asked about his bank accounts. The balances were not his. Somewhere inside Meta’s flagship AI assistant, wires had crossed — and a stranger’s private financial details had landed in the wrong conversation.

According to Reuters, that was not a one-off glitch. The news agency reviewed more than two dozen internal company posts in which Meta staff described Muse doing things nobody had asked for.

It answered questions that were never posed, surfaced orders nobody had placed, and reminded people about messages on apps it was never allowed to touch.

Meta calls the cases “isolated incidents”. The employees who lived through them call it something else — a warning about what happens when an AI agent is given the keys to your digital life before anyone has checked whether it knows whose life it is looking at.

What Went Wrong Inside Meta’s Muse AI Agent

Muse launched on September 8, 2026, as Meta’s big bet on personal AI. It is powered by the company’s Muse Spark model and runs inside a dedicated virtual machine Meta calls the Muse Secure VM.

The bet is working, at least on the numbers. Muse has passed 10 million downloads, and Meta says about a quarter of new users come back every day.

But inside the company, a feedback group where employees flag problems to the team building Muse has been filling up with reports of the agent going off-script.

Staff described receiving messages about childcare requests they had never made and career advice they had never sought.

Others reported product orders they had not placed and bank account balances that were not theirs. In one case, Muse booked a short-term rental stay without being asked to. The pattern, employees feared, was that the agent was mixing up conversations between different accounts — and exposing one person’s private data to another.

The Signal Message That Should Never Have Arrived

The most unsettling report involved an app Muse should have had no way of reading at all. An employee using the agent on her personal phone received a reminder to respond to a message from her sister.

The message had arrived on Signal, the end-to-end encrypted service that competes with Meta’s own WhatsApp.

“This seems like an extreme privacy violation. I definitely did not grant it access to other apps, or recall being asked about this,” she wrote in an internal post, according to Reuters.

Reuters did not stop at the employee’s account. Working with staff at Signal and researchers at the AI Now Institute, it ran its own tests to replicate the problem — and found the issue was ongoing, and could be affecting other people too.

How the Meta Muse AI Agent Could Read Messages It Was Never Given

So how does an AI agent end up reading messages from an app it was never connected to? The answer, according to the investigation, lies in the permissions Muse is granted when it is set up — permissions that many users may not realise they have given.

On phones running Google’s Android system, Muse is set up by default with access to push notifications, a colleague of the affected employee discovered after a similar surprise.

Notifications carry message previews in plain text, so an agent that can read them can effectively read your messages — even from encrypted apps — because the text has already been decrypted and displayed on your screen.

On Apple’s macOS, the route is different but the result is similar: accessibility features and screenshot tools could let the agent see what was on screen. The same request was blocked on an iPhone, and Apple has since restricted that pathway while adding more permission prompts for AI agents.

That is a problem that reaches far beyond one app. As Google’s own Gemini privacy reversal showed earlier this month, the AI tools now sitting closest to our inboxes and files can quietly gain access to far more than their users expect.

Meta Calls the Cases “Isolated Incidents”

Meta’s response has been to play the reports down. Daniel Roberts, a Meta spokesperson, told Reuters the examples were “isolated incidents” and said the company prioritises safety, tests extensively and keeps updating its products to improve protections.

Roberts also said Muse can access other apps only when the user has explicitly granted permission. Many of the messages employees believed were meant for someone else, he added, were later found to be AI hallucinations — invented details, not leaked data — which have since been addressed.

There is an important caveat in the reporting itself. Reuters said it could not confirm whether the bank balances shown to the employee were real data belonging to another person, or simply numbers the AI had made up. That distinction matters, and it is one reason the story deserves a careful read rather than a panic.

But the Signal case is harder to wave away. A reminder about a message on an encrypted app, in tests where Reuters found the behaviour still occurring, points at something more structural than a one-off hallucination.

The agent appears to be crossing boundaries between accounts and apps that should be sealed.

What Reuters Could Not Prove — and Why It Still Matters

The honest uncertainty at the heart of this story is worth stating plainly. Nobody has yet shown how many Muse users have been affected, or how often the mix-ups happen. The reports come from Meta employees testing the product, not from a public survey of its 10 million users.

Yet the pattern around the edges is hard to ignore. This is not the first time Muse’s boundaries have looked porous. In October, researchers found that share links generated by the agent could be discovered through passive DNS searches, exposing thousands of conversations to anyone who knew where to look.

That was a different failure mode — links leaking, rather than the agent itself confusing one user for another — but the shape of the problem is the same. Each time, the agent’s reach turned out to be wider than its users had been led to believe.

The AI security picture has been worsening all year; an earlier breach at Fireworks AI exposed customer datasets and model weights.

Why the Meta Muse AI Agent Story Matters for Every AI Assistant

Muse is not the only agent racing to take over your computer. Google has just given Gemini the ability to operate Mac apps and files, and Apple is reportedly building its own computer-controlling agent. The industry’s direction is clear: agents that can see your screen, read your notifications and act on your behalf.

That makes the Muse episode a test case for the whole category. If an agent cannot reliably tell one user’s data from another’s, or cannot tell which apps it is allowed to touch, then every permission it holds becomes a risk.

And the more useful the agent becomes, the bigger that risk grows.

The industry has started to build safeguards. OpenAI recently unveiled a control room that can halt a rogue model in 35 seconds. But safeguards only work if companies find the failures before their users do.

What Happens Next for Muse

For now, Meta says the reported issues have been addressed and that safety remains a priority.

Watch for three things in the coming weeks. First, whether Meta publishes a more detailed account of what went wrong and what it fixed.

Second, whether regulators or consumer watchdogs pick up the Signal findings, which raise questions that go beyond one company. And third, whether rival agents face the same scrutiny — because if Muse can read messages it was never given, the same permissions exist in every agent that asks to watch your screen.

The fishing forecast was harmless. The bank balances that followed it were not — whether they were real or invented, they showed an agent reaching for information it had no business holding.

As AI assistants move from answering questions to running our computers, that is the line that matters most. Muse’s early stumbles suggest the industry has not yet learned how to draw it.

Read the full investigation at Reuters.

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top