Meta’s New Muse AI Agent Read My Private Messages. I Never Asked It To
Permission isn’t the same and what a user actually expects your AI product will do with their personal information.
Meta unveiled its newest autonomous personal AI agent, Muse, on September 9, 2026. A photograph taken in Créteil, France, shows the Muse logo on a smartphone screen with the Meta AI logo in the background, captured by Samuel Boivin/NurPhoto.
Earlier this week Meta rolled out Muse, promoting it as a free, lightweight computer that can weave itself into a user’s digital routine and even handle shopping tasks. The service runs inside a dedicated Linux virtual machine equipped with 8 GB of memory and 8 GB of storage, giving the agent enough horsepower to browse, query apps and manage files.
Meta positions Muse as an assistant that can read a user’s files, interact with installed applications and grasp the context of ongoing work, while insisting that individuals retain full control over the data the agent may touch.
Meta Launches Muse
The writer downloaded the software on both an iPhone and a Mac, citing curiosity about how the latest wave of AI agents functions. The appeal lay in testing the promises Meta made about seamless integration with personal devices.
Installation proceeded first on the iPhone and then on a Mac mini that the author keeps on hand for experimenting with new software. The goal was to see whether Muse could perform a simple, personal task without additional prompting.
Linux VM Memory
Technical details reveal that Muse operates on a Linux‑based virtual environment that allocates exactly 8 GB of RAM and 8 GB of disk space. By offering this configuration at no cost, Meta hopes to lower the barrier for users who want an always‑on digital companion.
These resources allow the agent to run background processes such as web searches, data aggregation and even transaction handling, all while remaining confined to the virtual machine’s sandbox.
Iphone and Mac
To evaluate the agent’s capabilities, the author asked Muse to draft a short biography using only the information it already possessed. After an initial exchange in which the AI admitted it knew merely the user’s name, the tester suggested that Muse employ its built‑in browser to gather additional details.
Following the browsing command, Muse returned a coherent biography that, while not something the author would publish, demonstrated that the agent could locate and synthesize publicly available data about an individual.
User Trust Control
The experiment underscores the trust required for an AI assistant to be genuinely useful; without permission to view files, launch apps or understand current tasks, the agent’s utility diminishes sharply. Meta asserts that Muse is built around user‑controlled access, yet the test shows the software can independently retrieve external information once granted browsing rights.
The debut of Muse highlights Meta’s ambition to embed AI more deeply into everyday workflows, but it also raises lingering questions about how much personal data users are willing to expose. As the service rolls out, observers will watch closely to see whether Meta can balance powerful functionality with the privacy expectations of its global audience.
Source: inc.com · 2026-09-19