started as a weather question. An employee asked a chat assistant about the tide, and instead of answering, the assistant added a line nobody had requested.

“BTW, two bank updates, your Capital One balance is $2,722.58, your Charles Schwab balance is $5,899.57.”
The employee who posted it did not recognise those numbers. They belonged to someone else.
That is the moment Reuters pulled apart in its 18 November investigation into Meta’s AI agent Muse. The report, drawn from internal employee posts, describes a system that keeps mixing up conversations across different accounts, surfacing other people’s messages, purchases, and account balances in front of testers. Muse launched in September and has passed ten million downloads, which makes it one of the fastest-growing apps on record.
Meta has been building its AI agent strategy since early 2026, when it launched Muse as a chatbot that could hold conversations, summarize messages, and complete simple tasks across your apps. The September launch was treated as a consumer Play: a friendly, emoji-driven assistant that could message your friends, order groceries, and set reminders. The 10 million download milestone in two months was the proof point — people wanted an agent that could act, not just talk. But the Reuters reporting shows the hard part nobody fully solved. An agent that can reach into your email, your bank, and your messages is an agent that can leak other people’s data. The same capability that makes Muse useful is what makes it dangerous. Meta’s internal testing surfaced the problem before the public ever used the app — but the question is whether the fix holds when the agent’s reach expands. And expand it will. The $300 Charm, with its always-on microphone and on-device intelligence, is the next step in that direction.
The leaks are a preview of what happens when you hand a machine an agent that can reach into your life — email, banking, messaging, photos — and act on your behalf. The headline of 2026 so far has been computer use: Google giving Gemini the ability to click and type on your Mac, Claude and Codex running tasks across your desktop. The flip side, the part nobody pays for in the cheerful headlines, is what happens when the agent’s boundaries leak.
Two examples from the Reuters piece stick out. One employee asked Muse about weather and tide. Muse replied, unprompted, with the bank balances of someone else. In another case, an employee on their personal phone got a Muse reminder to reply to a message from their sister on Signal, the end-to-end encrypted messaging app. The employee wrote: “This seems like an extreme privacy violation. I definitely did not grant it access to other apps, or recall being asked about this.”
Reuters does not say what caused the mix-ups, and the reporting is careful about what it does not know. What is clear is where the data has to travel. A language model can only repeat what the software around it has placed into its context: stored memories, past conversations, data pulled from connected apps. For someone else’s balance to appear in a chat, something in that chain fetched context under the wrong identity, or read from a store more than one account could reach. The data arrived in fluent prose, which is exactly what made it hard to spot.
What is actually happening under the hood? AI agents don’t store everything in a single, easy-to-control database. They read from a chain of context sources: the model’s own training, the user’s memories, conversation history, and data pulled from connected apps. Each retrieval is a separate decision about whose data is visible. If that retrieval isn’t strictly scoped to the signed-in account, the model sees another person’s balance. The Reuters reporting suggests this is a retrieval scoping problem, not a model failure. The model did exactly what it was told; it just saw data it shouldn’t have had access to. That’s why the fix isn’t ‘make the model smarter’ — it’s ‘make the retrieval layer respect account boundaries.’ It’s a plumbing problem, not a brain problem.
Meta would not go further than its statement to TechCrunch. Employee testing is core to how the company develops products, the spokesperson said. Employees get different versions of Muse than the public, and the testing exists to find safety and privacy problems before features ship. That is a fair defence at the testing stage. It does not answer the question the reporting raises: whether the public version has the same gaps, whether any public user was affected, and whether the problems were fixed at all.
The cheaper way to read this is as a contained incident. The harder reading is that it is a design problem, and it is about to get more visible. Meta is expected to ship the Muse Charm, a keychain-sized device with a screen and a fingerprint sensor that lets you talk to Muse without unlocking your phone or opening an app. The Information reports the price at $300, with up to a year of free Verizon 5G data bundled in. A $300 hardware gate to an agent that already has trouble keeping one person’s data separate from another’s is not a configuration the privacy community will ignore.
The privacy community’s reaction will be shaped by regulators. The EU’s GDPR framework already has the language for this: a data isolation failure is a personal data breach, full stop, and the ’employee testing is fine’ defence only holds if the testing is genuinely limited, properly disclosed, and corrected before public release. UK regulators, the US Federal Trade Commission, and — closer to home for many readers — data protection authorities in Southeast Asia are watching. Meta has not released a formal incident summary or confirmed whether the public app has the same flaws. If it doesn’t, or if the Charm’s always-on microphone and always-online connectivity broaden the attack surface, the question moves from ‘did Meta fix it?’ to ‘who is liable when an AI agent makes a mistake?’
This is not academic for me. I manage ICT systems, I run AI-assisted workflows, and I have learned that the moment a tool can reach into your data — messages, files, accounts — it stops being a convenience feature and becomes a security boundary. I’ve seen what happens when a team treats an API as "just for read-only" and then discovers it can write too. The Muse reporting points at a specific failure: context retrieval that is not scoped to the signed-in user. It’s the same class of bug as a shared cache in a multi-tenant database — someone else’s rows show up in your query results. The fix is not a patch of the model. It is a discipline in how agents read: each memory lookup, each connector call, keyed to the user in front of the screen, with no shared cache that can return another user’s rows. Meta’s own employees found it in internal testing. The question is whether they carry that discipline into the public version, and into the $300 Charm.
For developers and ICT teams, the practical takeaway is a checklist, not a panic. Seed a handful of test accounts with unique marker data, run many sessions at once, and fail the build if any account’s markers ever appear in another account’s replies or logs. Record for every response which sources contributed to its context, so a leaked balance can be traced back to the exact retrieval step. And treat consumer AI agents as a named category in vendor risk tracking, not an afterthought tucked into the general AI policy. TheSignal reminder is the tell: an agent should be able to show each user which apps and data it can read, and most cannot.
Two other Bleuken posts frame this best. Your Agent Will Pick the Hackable One argues that the security boundary is the install, not the model. How to Vet an AI Coding Agent Before Trusting It With Your Code walks through the practical vetting steps, which apply just as well to any agent that can read your data. And if you want the retention angle, Google Keeps Gemini Chats for 2 Months, Even With History Off shows how much of our data lives in systems we assume are off. Together they make the case that the AI industry has built capability faster than the boundaries around it.
The agentic economy is moving fast. Meta wants to put an AI agent in your pocket the size of a keychain, Google wants one that can click and type for you, and every major cloud provider is racing to build the platform where agents live and work. Neither is going to stop. But the Muse story is a reminder that the safest agent is the one you can keep out of the room. You can’t un-invent an agent that can act for you. Whether Meta can build that, with a $300 Charm around your neck or an app on your phone, is the question the next few months will answer. The rest of us can only do the honest thing: ask what it means before we let it in.