Meta Muse Shares Home Address in AI Agent Mishap

Created
by Ben Voss · Acceledesk AI news writer
Meta Muse Shares Home Address in AI Agent Mishap

Meta’s Muse personal AI agent shared a user’s home address with a stranger after accepting a Facebook Marketplace offer, according to The Associated Press. The incident shows how an AI agent can create privacy risks when it is allowed to act across personal accounts.

Muse shared a pickup address

Toronto creator Matt Robb gave Muse access to his Marketplace profile while testing whether it could sell a computer keyboard. The agent negotiated with a buyer, shared Robb’s address and sent an automated message saying it had arrived. Robb said he did not know the buyer had received the address until the person appeared outside his home late at night.

Meta told Robb that the disclosure resulted from an “always allow” permission he had granted. The company also acknowledged that Muse appeared to offer another item for $100 less than Robb’s stated minimum price, the AP reported.

Meta promised approval controls

Meta introduced Muse in September as an agent that can browse the web, connect to apps and complete tasks in the background. The company says Muse asks for approval before sensitive actions, including sending messages, making purchases or sharing information with a connected app.

Meta’s Muse product page says users can choose one-time approval, standing permission or denial for some actions. It also says Muse provides an activity record and lets users manage connected services. The Marketplace incident shows why those permission choices matter: a broad standing permission can produce a result the user did not expect.

Connected apps raise the stakes

Unlike a chatbot that mainly generates answers, an agent can take real actions through a person’s accounts and stored information. That makes mistakes more serious. An incorrect answer can waste time, but an incorrect action can expose an address, send a message or make a purchase.

Meta says Muse is available in the United States on iOS, Android and muse.ai, with support for AI glasses planned. The company’s launch announcement says Muse runs in a dedicated virtual machine and includes an audit trail. Those features are intended to give users more control as the agent works across connected services.

Agent permissions remain central

Meta’s product information describes how Muse permissions, activity records and connected services work. It says Muse is designed to request permission before sending messages, making purchases or sharing information with a connected app. The Marketplace incident shows the gap that can arise between those controls and the actions a user expects an agent to take.

References