Meta sells Muse as an agent that handles everyday errands. Karl Bode’s Techdirt critique assembles a less comforting launch record: a reported Mac spying vulnerability, a user’s Marketplace agent selling too cheaply and disclosing his address, and a report that the agent referenced private Messages without an explicit grant. Wired also found Muse building profiles of users’ contacts and collaborators; 404 Media reported a scramble to fix security holes before launch.
The incidents have different causes. A vulnerability is not the same as a user misunderstanding permissions, and an engineer’s warning about a future breach is not a confirmed breach. But they share a design problem: the more accounts, messages and actions an assistant can reach, the more expensive every boundary failure becomes—for the user and for people who never chose the product. A cheerful avatar does not narrow the access.