Meta’s AI agent Muse is facing questions about how it separates personal information after employees reported seeing material that appeared to belong to other accounts. TechCrunch, citing a Reuters report based on internal company posts, said the incidents occurred during employee testing. The available reporting does not establish that the same behavior has occurred in the public version of Muse.
Key points
- Employees said Muse surfaced content they did not recognize, including childcare requests, career advice, product orders, and bank account balances.
- In one account, an employee said Muse added two bank balances to a conversation about weather and tides. Another employee said Muse prompted her to respond to a message from her sister on Signal, despite not recalling granting the agent access to other apps.
- Meta said employees have access to different versions of Muse from the public and may be testing features or behavior that are not available outside the company. The company described employee feedback as part of product development and a way to improve safety and privacy protections.
The concern is more serious than an ordinary incorrect answer. If an agent displays information that belongs to another account, it raises questions about account separation, permission boundaries, and whether users can understand where a piece of information came from. These issues matter especially for products designed to connect conversations with personal services and reminders. Users need clear controls over what an agent can access, and systems need reliable ways to prevent information from one account appearing in another.
At the same time, the reported examples come from internal testing and employee posts. They do not, on their own, establish the technical cause, how widespread the behavior was, or whether public users were affected. Meta’s response highlights the difference between internal and public builds, but the material provided does not explain what triggered the mix-ups or whether the reported issues have been fixed.
The report says Muse launched in September and has surpassed 10 million downloads. As agent-based products reach more people, confidence will depend not only on their usefulness but also on predictable handling of personal data. Employee testing can help identify failures before release, but it needs to examine not just whether features work, but whether permissions are understandable and account boundaries hold under real-world use. Until the company provides more detail, users should treat unexpected sensitive information displayed by an agent with caution.
Source: Web Search
Comments
Checking sign-in status...
Loading comments...