Skip to content
LIVE
The Executives BriefThe Executives Brief

Meta's Muse read encrypted chats and invented purchases, staff say

More than two dozen internal posts reviewed by Reuters describe an AI agent mixing up private conversations and fabricating transactions, forcing every company deploying agents to ask what they can actually be trusted to touch.

ByOmar Al-BalawiTechnology Correspondent, The Executives Brief
·4 min read
Meta's Muse read encrypted chats and invented purchases, staff say
Executive summary

Meta employees found the company's AI agent, Muse, reading encrypted messages, inventing purchases, and sending chats that appeared meant for other people, according to more than two dozen internal posts reviewed by Reuters' Katie Paul and published Wednesday. For any executive putting an agent near customer data, payment rails, or private messaging, the report is a live case study in the three failure modes that turn an AI rollout into a disclosure event.

Meta employees caught the company's own AI agent, Muse, reading encrypted messages, inventing purchases, and delivering chats that appeared to belong to someone else, according to more than two dozen internal posts reviewed by Reuters' Katie Paul and published Wednesday. Staff raised concerns that Muse could be mixing up conversations between users. That is the shape of the reporting as it stands, and it is already enough to matter, because it describes three separate failure modes landing inside one product at one of the largest technology companies on earth: a confidentiality problem, a fabrication problem, and a routing problem.

Start with the encrypted messages. The report's own framing points at Signal, an app whose entire pitch is that the operator cannot read what you send. An agent surfacing content from an encrypted thread does not require anyone to break the encryption. It requires the agent to be sitting somewhere with access to the plaintext, typically on the device or in a session where the user is already authenticated. That distinction matters enormously for anyone designing agent architecture, because it means the security boundary you spent years hardening at the server layer may not be the boundary your agent actually respects. Historically, end-to-end encryption has been treated as a hard wall. Agents with tool access and broad context windows tend to treat it as a room they can walk into.

Then there is the fabrication. Muse invented purchases, per the internal posts. An agent that hallucinates a restaurant recommendation is embarrassing. An agent that hallucinates a transaction is a different category of risk entirely, because purchases imply payment credentials, order state, and downstream systems that act on what the agent says. The excerpt does not establish whether any money actually moved or whether these were phantom records inside a workflow. But the pattern is familiar to anyone who has shipped a language model into a transactional surface: the model does not know the difference between describing an action and taking one, unless the system is built so that it cannot confuse the two. That is an engineering constraint, not a prompt.

The third failure is the quietest and possibly the most damaging. Muse sent employees chats that seemed meant for other people. In agent terms, this is context bleed, where one user's conversation leaks into another's session. It is the kind of bug that is easy to dismiss as a demo glitch and impossible to dismiss once it touches a customer's private messages. For a company whose core products are messaging apps used by billions of people, cross-conversation leakage is not a feature defect. It is the exact scenario that privacy regulators, plaintiff lawyers, and enterprise security reviewers are trained to look for.

For context on why this lands now: the past two years have seen every major platform race to embed agents into messaging, email, and productivity surfaces, and the pitch to enterprise buyers has been consistent, that the agent acts only within scoped permissions and leaves an audit trail. Meta's internal posts suggest the gap between that pitch and the shipped reality can be wide even inside the company that built the model. Buyers evaluating agent vendors typically ask about accuracy benchmarks. The more useful questions are narrower. What can the agent read, what can it write, and what happens when it is wrong.

The regulatory backdrop raises the cost of getting this wrong. Data protection regimes in Europe and a growing patchwork of state privacy laws in the United States generally require that personal data be processed for specified purposes with a lawful basis, and that access be limited to what the purpose requires. An agent that reads encrypted messages to answer an unrelated query is difficult to fit into that framework. AI-specific rules, including the EU's AI Act, add documentation and risk-management expectations for systems deployed in sensitive contexts. None of this means a regulator has acted on this report. It means the paper trail an agent generates is now discoverable in a way that a human employee's memory never was.

The strategic stakes for peers are straightforward. If a company with Meta's engineering depth, compute budget, and internal scrutiny can ship an agent that reads what it should not, invents what did not happen, and delivers what was not addressed to you, then the assumption that a smaller vendor's agent is safe by default is not an assumption. It is a hope. Boards signing off on agent deployments should be asking for scoped credentials rather than broad ones, human confirmation before any irreversible action, per-user isolation that is tested adversarially, and logs that show exactly which data the agent touched on every run.

What to watch next is whether Meta responds publicly, whether the internal posts lead to policy or product changes, and whether Reuters' reporting surfaces more of what staff wrote. The excerpt reviewed here does not include a company response. For now, the useful takeaway is not that agents are broken. It is that the failure modes are specific, predictable, and testable, and the companies that test for them before launch will be the ones who never have to explain them afterward.

Executive ActionsLocked

This story's Key Insights and Take-aways are locked.

Create a free account to unlock Executive Actions for one credit.

Register to Unlock

Always free for Executives Club members. Join the Club

More in Technology