Key Notes
- Meta says many apparent Muse conversation mix-ups were hallucinations, while Reuters could not verify a cross-account financial-data leak.
- Reports of unexpected private-message access put the scope of device permissions and user consent under scrutiny.
- Meta says employee test builds can differ from the public product, and its published architecture includes separate controls for external actions.
Meta is defending how its Muse AI agent handles personal information after employees reported unexpected access to private messages and conversations they did not recognize. The company says many apparent account mix-ups were hallucinations, leaving a critical distinction between an agent inventing sensitive details and actually exposing someone else’s data.
In a November 19 response, Meta told TechCrunch that employee test builds can differ from the public product. That qualification matters when assessing reports from internal testing, although it does not resolve the broader question of what people understand they have authorized an agent to read.
Private Messages and Unexpected Replies
A Reuters investigation reported on November 18 that staff had raised concerns about private-message access, unsolicited purchase-related conversations and apparent financial information they did not recognize. Reuters could not establish whether the purported bank balances belonged to other people or were invented.
In a November 19 explanation of the investigation, the reporter described an employee whose agent referred to a family member’s Signal message despite the employee not intentionally connecting that service. Meta maintains that Muse accesses information through permissions users have granted.
The disagreement concerns the reach of those permissions and the actions users reasonably expect them to enable. Access to readable messages on a device raises a different security question from defeating the encryption protecting messages in transit. The reporting does not establish that Muse broke Signal’s encryption.
Hallucinations Complicate the Privacy Picture
Meta has attributed many apparent conversation mix-ups to fabricated responses and pointed to a previously fixed bug. Reuters’ reporter said employees continued discussing unexpected messages from September into November. Those reports establish concern about the agent’s behavior; they do not, by themselves, prove a cross-account data breach.
For a personal assistant, either explanation creates a trust problem. An invented financial update can still mislead the recipient, while genuine access outside a person’s expectations would call the consent process into question. Users need a dependable way to understand what information an agent retrieved and which actions it actually performed.
The Safety Design Faces a Practical Test
Meta’s published safety architecture describes a separate Sentinel system that authorizes connector actions and network access. It also separates sensitive credentials from the main agent’s runtime. These are safeguards intended to limit what the assistant can do, rather than a guarantee that it will never make mistakes.
The company acknowledges that agents can err and encounter attacks through material they read. The practical test is whether technical protections, permission prompts and visible activity records give people an accurate understanding of the access they are granting.
The scrutiny arrives as Muse expands into business calls and custom workspaces. The more jobs an assistant takes on, the more consequential the boundary between useful context and unwanted access becomes.
Disclaimer: AIstify is an independent media brand owned and operated by NuvexMedia LLC, publishing news, research, and insights on artificial intelligence, emerging technologies, automation, and related industries. NuvexMedia LLC invests in and collaborates with companies across the AI, technology, software, and digital innovation sectors. These relationships do not influence AIstify’s editorial coverage, and the publication maintains full editorial independence to provide accurate, timely, and objective information. © 2026 NuvexMedia LLC. All rights reserved. This content is for informational purposes only and should not be considered legal, tax, investment, financial, or other professional advice.