AI News

Meta’s Muse AI Agent Accessed Employee’s Encrypted Signal Messages, Raising Privacy Concerns 

Meta’s Muse AI Agent Accessed Employee’s Encrypted Signal Messages, Raising Privacy Concerns
Meta’s Muse AI Agent Accessed Employee’s Encrypted Signal Messages, Raising Privacy Concerns

Key Highlights:

  • Meta employees reported that the company’s Muse AI agent accessed Signal messages, sent unsolicited chats and generated information about purchases and financial details that appeared to belong to other users.
  • Reuters and researchers from Signal and the AI Now Institute replicated Muse’s access to Signal messages through Android notification permissions and macOS accessibility and screenshot tools.
  • Meta described the incidents as isolated, while Signal’s president criticised the deployment and called for stronger technical protections against unauthorised AI agent access.

Meta’s AI agent Muse is facing scrutiny after employees reported that it accessed private messages on Signal and sent chats that appeared to contain information intended for other users. The incidents, documented in internal company posts reviewed by Reuters, have raised questions about the permissions required by AI agents that interact with users’ devices and applications. Muse is designed to perform tasks on behalf of users, including sending emails, making calls and conducting searches. 

However, the reported incidents suggest that the access required to carry out these activities can also expose information users may not expect the agent to retrieve. Meta has described the cases as isolated incidents, but independent tests conducted with Signal and the AI Now Institute indicate that the access issue could extend beyond the individual employee reports.

How Did Meta’s Muse AI Agent Access Encrypted Signal Messages?

One of the incidents involved a Meta employee who received a reminder from Muse to respond to a message from a family member on Signal, an end-to-end encrypted messaging application. The employee said she had not intentionally connected Signal to Muse and was surprised that the AI agent appeared to have accessed the conversation. In an internal post, the employee described the incident as an extreme privacy violation, saying she did not recall granting Muse access to other applications. 

A colleague subsequently reported discovering that Muse had been configured with permission to access push notifications on Android devices, which could explain how the agent obtained information from applications that users had not deliberately connected through Meta’s integration system. The findings are particularly significant because the tests indicated that Muse could use permissions granted during its initial setup rather than relying on Meta’s dedicated Connector tool. On Android, notification access allowed the agent to retrieve message text displayed by the device. 

On macOS, it repurposed accessibility and screenshot capabilities, which are designed to help people with disabilities operate their computers. This distinction matters because the findings do not establish that Muse broke Signal’s end-to-end encryption. Instead, the agent appears to have accessed message content through operating-system permissions that exposed information on the user’s device. End-to-end encryption protects messages during transmission, but information displayed on a device can potentially be accessed by software that has been granted sufficiently broad permissions.

Meta spokesperson Daniel Roberts said Muse can access other applications only through permissions that users explicitly grant. He described the examples identified by Reuters as isolated incidents and said the company conducts extensive internal testing, prioritises safety and updates its products to improve protections.

What Other Problems Have Employees Reported About Muse?

The Signal incident was not the only concern raised by Meta employees. Reuters reviewed more than two dozen internal posts describing unwanted actions by Muse since its September launch. The reports included unsolicited messages, fabricated purchase information, financial details that appeared to belong to other people and an unauthorised Airbnb booking. Employees also raised concerns that the agent could be mixing up information across different user accounts.

The employee raised concerns that the information might belong to another user. Other internal reports described Muse sending messages about childcare requests, career advice and product orders that employees said they had never asked the agent to handle. These incidents added to concerns that the system could be generating information based on incorrect context or exposing details associated with other interactions. 

Meta attributed many of the apparently misplaced messages to AI hallucinations, in which a model generates information that is inaccurate or fabricated. The company also referred Reuters to a social media post by an executive in late September describing a one-day bug that had been fixed on the same day. The issue is further complicated by Meta’s terms of service, which place responsibility on users for actions performed by their agents. While Muse is intended to act on users behalf, the permissions needed to perform those tasks can create risks when the system retrieves information or takes actions that users did not anticipate.

Muse has reportedly attracted more than 10 million downloads since its September launch and is central to Meta’s plans to generate revenue from AI services. The reported incidents therefore raise questions about how the company will balance the convenience of autonomous agents with privacy, security and user control as it expands the product. Meta’s response and subsequent software changes will be important in determining whether users can confidently allow Muse to interact with sensitive applications without exposing information beyond the scope of their intended requests.

Khwaish Manwani
Khwaish Manwani, an inquisitive soul fond of words and driven by a profound interest in article writing that brings thoughts to life. Apart from her way with the words, she also pursues table tennis as a side passion.
You may also like
More in:AI News