The most secure messenger couldn't protect AI.

✨ Megiddo

✨ President ✨
Staff member
Administrator
Messages
4,022
Likes
1,873
Points
2,130
Muse was found to have information about bank accounts, family affairs, and other people's conversations.

image


Meta's Muse* AI agent, with over 10 million downloads, learned to read Signal messages without the permission users expected. Meanwhile, even more disturbing behavior was discovered within the company. The assistant sent seemingly unrelated messages, displayed bank balances, and made unwanted purchases.

On November 18, 2026, journalists uncovered the contents of more than two dozen internal messages from Meta employees complaining about Muse's behavior. The app was released in September and quickly became one of the most popular AI services. The agent can manage email, calendars, and shopping carts, autonomously performing tasks through connected apps.

Employees were particularly alarmed by messages that appeared to be from other people's accounts. One user received notifications about Capital One balances of $2,722.58 and Charles Schwab balances of $5,899.57, despite never having linked bank accounts. Others received career advice, order details, and reminders about undisclosed family matters.

The most telling case involved a Meta employee. Muse unexpectedly reminded her sister to reply to her in Signal, even though the woman had not authorized the agent to read the conversation. Signal and AI Now Institute specialists helped investigate the complaint. Experiments confirmed that Muse was capable of obtaining the contents of secure messages by bypassing the Connector app's native connection mechanism.

On Android, the agent read the text through authorized push notification access. On macOS, it used accessibility and screenshot tools. In both cases, Muse invoked previously granted system privileges, although the user might have been unexpected. On iPhone, similar attempts were blocked.

Remarkably, Signal's encryption is not compromised by such actions. The agent receives the information only after the message is decrypted on the device. After discovering the issue, Apple limited the use of accessibility tools in macOS and introduced additional permissions for AI agents.

Muse's unpredictability also affected financial transactions. Employees reported fictitious purchases, incorrect prices, and booking non-refundable Airbnb accommodations without permission. Meta representatives described the reported cases as isolated, and attributed some of the suspicious messages to the neural network's hallucinations. Journalists were unable to confirm a genuine leak of banking information between accounts.

The investigation also states that the developers attribute the strange behavior to the model itself and the software components of the agent. Signal President Meredith Whittaker criticized Meta's approach to privacy and proposed giving apps the technical ability to block access to AI agents.

Muse has encountered problems before. In early November, Meta disclosed vulnerabilities that allowed it to obtain messages and tokens from employees testing the agent. The current cases highlight a different danger. Even without hacking, the AI assistant can extend its intended access beyond the user's intended scope if system permissions allow it to read other people's apps.

* Meta and its products (including Instagram, Facebook, and Threads) are considered extremist and are banned in Russia.