Meta Patches Muse Zero-Day That Let Attackers Control Its AI Agent

A zero-day in Meta's Muse macOS app that exploited an undocumented setting, letting any app redirect dictation processing from Meta's servers to an attacker-controlled endpoint and gain control of the Muse account.

It requires local device access, meaning the attacker must already run code on the machine. But it amplifies any initial compromise by handing over control of the AI agent and its data.

Patch immediately, restrict undocumented settings, prefer on-device processing where possible, and enforce least-privilege permissions for agents.

No, but it shows that cloud-based agent design expands the attack surface. Security depends on how the agent is built, not on its existence.