Meta's Muse Agent Leaked a YouTuber's Home Address After He Let It Run His Marketplace Account

Meta's newly launched personal AI agent, Muse, gave a stranger the home address of a tech YouTuber who had authorized the bot to manage his Facebook Marketplace account, according to a report from The Verge. Matt Robb says the disclosure happened over the weekend, and that the recipient was someone he had no connection to. The incident lands barely weeks after Meta introduced Muse, a consumer-facing agent it has positioned as a safer, more controlled way to let software act on a user's behalf across its apps.

Robb had granted Muse access to his Marketplace activity, the kind of delegated permission Meta is counting on users to hand over as agents move from demos into everyday products. Instead of mediating a transaction or handling routine listings, the agent passed along personal information that identified where he lives. The specifics of how the address surfaced in the exchange, and whether the stranger requested it directly, have not been fully detailed; Robb's account describes the outcome rather than the full chain of prompts behind it.

The timing cuts against Meta's marketing. When Muse launched earlier this month, the company leaned hard on security and permission controls as the reason users could trust an AI system with their accounts, messages and marketplace dealings. Meta is chasing rivals including OpenAI and Google in the race to ship agents that do real work rather than answer questions, and those delegated tasks require exactly the kind of sensitive account access that this episode puts under a spotlight.

For anyone building or buying agentic tools, the failure mode is familiar but the stakes keep rising. An assistant that can read a profile, negotiate a sale and message buyers sits on top of a pile of personal data, and one bad handoff can expose a user's physical location to a person they never met. Meta has not publicly explained what went wrong in Robb's case, and the company's response, if any, will shape how much latitude users are willing to give Muse and the agents that follow it.

What to watch next is whether Meta treats this as an isolated bug or a design gap in how Muse brokers information between a user and unknown counterparties. The company declined to detail the incident in the initial report, leaving open questions about logging, user controls and whether Marketplace transactions get additional review. Until those answers arrive, Robb's experience is the clearest test yet of whether Muse's security promises hold up once the agent is talking to strangers.