A critical zero-day vulnerability has been identified in Meta's recently launched AI assistant, Muse, undermining the platform's security claims. Despite its design to handle sensitive tasks like making purchases and accessing WhatsApp or email accounts, the macOS application contains a flaw that allows locally executed code or terminal commands to bypass operating system-level protections. The vulnerability stems from an architectural design choice where any local process, regardless of its assigned macOS permissions, can modify a series of undocumented settings within the assistant. While many of these settings are minor, a specific flaw allows an attacker to redirect the transcription endpoint from Meta's secure servers to a malicious external address. By successfully hijacking this endpoint, an attacker can intercept the authentication token used for the user's Muse account. Once this token is compromised, the attacker gains complete, unauthorized control over the assistant's capabilities, including its access to the user's calendar, location, microphone, camera, and various third-party services. This discovery has prompted Amazon to begin blocking the Muse service. The vulnerability highlights a significant breakdown in the security model intended to prevent local applications from accessing restricted device resources and sensitive user credentials.