Article
AI News Privacy & Data AI assistant privacy

Meta's Muse AI agent under scrutiny after surfacing other people's data

Reuters found more than two dozen internal employee reports of Meta's Muse assistant surfacing financial data and messages it apparently had no authorization to see.

by TechDefused · Edited by IL
A person holds a tablet displaying a logo related to artificial intelligence. The background features a blurred setting with a coffee mug and headphones, suggesting a casual workspace environment. — Credit: Photo by Jo Lin on Unsplash c Photo by Jo Lin on Unsplash

Meta's Muse AI agent is facing scrutiny after internal employee reports reviewed by Reuters showed it surfacing information that appeared to belong to other people.

Muse, launched in September 2026, is designed to act across a user's apps and communications, carrying out tasks rather than just answering questions.

That makes misdirected data far more serious than a typical chatbot error.

In one case, an employee asked Muse about fishing weather and tides and received two bank balances he had not asked for: "your Capital One balance is $2722.58, your Charles Schwab balance is $5899.57".

He said he had never shared banking details with the assistant.

Another employee found that Muse had referred to a private Signal message from her sister, even though she had not knowingly connected the encrypted messaging app to the assistant.

Reuters worked with Signal and the AI Now Institute, a research group, to reproduce the behaviour.

They found Muse could reach Signal messages through Android notification permissions and through macOS accessibility and screenshot tools, bypassing its own app connector.

Similar attempts failed on an iPhone.

The wider complaint log included childcare reminders nobody had asked for, shopping orders, and conversations employees said they had never started.

Reuters could not confirm whether the data came from other users' accounts or was made up by the model.

Meta spokesperson Daniel Roberts called the cases "isolated incidents".

He said many of the apparent account mix-ups were likely hallucinations, where an AI model invents information and presents it as fact.

Apple has since tightened the macOS permissions that govern AI agents' access to personal data.

Muse has been downloaded more than 10 million times.

Read more about Meta Platforms

by TechDefused