Security researcher Patrick Wardle discovered a critical zero-day vulnerability in Meta’s new Muse AI agent for macOS that allows attackers to hijack user accounts and access sensitive data.
Key Points
- The vulnerability allows local apps or Terminal commands to modify undocumented Muse settings without requiring special macOS permissions.
- Attackers can redirect dictated voice prompts to unauthorized servers, enabling the theft of authentication tokens to control the AI agent.
- Exploitation can lead to malicious file creation, unauthorized camera access, and the retrieval of location data from linked iPhones.
- The flaw stems from Meta’s decision to process voice data on external servers rather than using Apple’s secure on-device dictation tools.
- Meta has reportedly released a software update to address the security gap, and users are urged to install the latest version immediately.