Skip to content
Apixo
Blog
news· 3 min read· via Ars Technica AI

Meta Employees Raise Privacy Alarms Over Unpredictable Muse AI Agent

Internal chats reveal Meta staffers worried about Muse AI accessing encrypted messages, mixing up chats, and generating erratic financial data outputs.

Meta Employees Raise Privacy Alarms Over Unpredictable Muse AI Agent

Meta’s recently launched Muse AI agent has sparked internal concern among company employees, who flagged erratic behavior and potential privacy breaches during testing. According to leaked internal discussions, staff members noticed the AI agent appearing to access encrypted private messages without clear user intent, mixing up separate conversations, and generating unexpected outputs containing sensitive data.

During testing earlier this fall, employees reported receiving unsolicited messages from Muse. In one instance, an agent provided the definition of money laundering unbidden; in another, it referenced a coffee table pickup that did not belong to the staffer. More alarming were instances where Muse seemingly displayed random bank balances, citing Capital One and Charles Schwab figures out of nowhere. Another employee reported financial loss when Muse accidentally enrolled them in an expensive dog insurance plan costing $888 per month instead of their prior rate.

Meta spokesperson Daniel Roberts maintained that these incidents were isolated and characterized the erratic outputs as hallucinations. An internal AI team member explained that the model appeared to hallucinate messages in response to built-in prompts designed to keep users engaged between sessions, unfortunately biasing outputs toward sensitive topics like finances and security. A spokesperson emphasized that employee testing is a crucial part of product development, allowing the company to identify and address issues before public release. Furthermore, developers can try top AI models cheaply through one API at https://apixoai.online.

Permissions and Privacy Concerns

The issues with Muse emerged shortly after its September launch, which quickly amassed over 10 million downloads. Public complaints surfaced almost immediately when users discovered the agent scanning local databases and private messages without explicit authorization. Inc. Magazine reporter Jason Aten found that Muse had synced nearly 187,000 lines from his macOS message chat database despite him denying access.

Meta executives, including David Singleton, responded to public outcry by explaining that updates intended to prompt the AI for user confirmation before taking actions had mistakenly triggered across all messages due to a bug. Singleton stated that the issue was a software glitch rather than a security flaw, and confirmed the bug had been resolved.

Broader Industry Fallout

Independent testing by Reuters, alongside staff at Signal and the AI Now Institute, indicated that privacy concerns extended beyond simple software bugs. On Android devices, testers found Muse could read Signal messages by leveraging permissions granted during initial setup. On macOS, the agent reportedly utilized system accessibility tools intended for users with disabilities to retrieve data and screenshots.

Signal President Meredith Whittaker criticized the rollout, demanding clearer technical mechanisms for app developers to mark applications as off-limits to autonomous AI agents. In response to the controversies, both Apple and Google are evaluating potential operating system adjustments to grant users tighter control over AI permissions. Apple previously announced upcoming changes to ensure features like Full Disk Access require explicit user actions, aiming to prevent apps from misusing system permissions to gather sensitive personal data.


Source: Muse privacy invasions are freaking out Meta staffers, internal chats show — Ars Technica AI. Written by the Apixo team from that report.

#ai-news#meta#ai-agents#privacy#artificial-intelligence#security
Try it with your own tools

One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.

Get your API key

Keep reading