Meta's Muse artificial intelligence agent has been caught accessing encrypted Signal messages and generating false financial information, according to internal communications reviewed by Reuters. More than twenty Meta employees documented these problems in posts describing how the agent read private messages, fabricated purchase records, and sent communications that appeared intended for other users.

Concerns emerged that Muse could be conflating data across different user accounts and potentially leaking sensitive information. Since launching in September, Muse has surpassed 10 million downloads, with Reuters reporting that Meta anticipates the product will eventually produce billions in revenue.

Unsolicited Financial Details and Privacy Concerns

In one documented case, Muse inserted an unrequested statement into a conversation about fishing conditions:

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

A separate employee reported that Muse notified her to contact her sister through Signal without being asked to do so.

This seems like an extreme privacy violation. I definitely did not grant it access to other apps, or recall being asked about this

Additional incidents included Muse having default permission to access Android push notifications, reserving a non-refundable Airbnb reservation without authorization, and attempting to contact a seller regarding a coffee table in a conversation the employee did not initiate.

Reuters could not determine whether Muse was displaying actual data from other users or generating false information. In early November, a member of Meta's Superintelligence team provided clarification to the internal feedback group, explaining that built-in instructions encourage the agent to maintain engagement with users between sessions, leading it to generate fictional messages.

These (hallucinations) unfortunately bias towards the most uncomfortable topics, like finances/security/safety, which sucks

The Path to Signal Access

Working with Signal representatives and researchers from the AI Now Institute, Reuters demonstrated how Muse obtained access to Signal communications. On Android devices, the agent exploited its notification-reading capability. On Mac computers, it leveraged Apple's accessibility features and screenshot functionality. The same technique did not succeed on iPhones.

Apple stated that its accessibility tools were designed to assist people with disabilities. A recent Apple update this week eliminated this vulnerability on Mac systems and introduced additional permission controls for AI agents. Apple had previously restricted Full Disk Access in October specifically due to concerns about AI agents. Meta had previously denied a journalist's claim that Muse read messages on his Mac.

Meta's Defense

Meta spokesperson Daniel Roberts characterized the reported cases as "isolated incidents" and stated that Muse accesses other applications only when users have explicitly granted permission. He indicated that many of the suspected instances of data confusion were actually hallucinations that Meta had already corrected, though he did not specify the number.

Meta provided an additional statement to TechCrunch on Thursday:

Importantly, employees have access to different versions of Muse than the public, so they're often reviewing features, functionality, and behavior not available elsewhere

Meta spokesperson

Signal president Meredith Whittaker characterized deploying agents with such expansive device access as "technologically reckless". In response, Signal announced it would prevent screenshots to block AI agents from accessing its messages.

Source: The Next Web