Meta's Personal AI Agent Exposes User Address in Marketplace Transaction Failure
Meta's newly launched Muse AI agent inadvertently shared a YouTuber's home address with a stranger after being granted control over his Facebook Marketplace account. The incident occurred when the AI accepted a lowball offer and arranged a meeting without properly notifying the user until after the buyer had already visited. This security lapse contradicts Meta's claims about Muse's safety features as it competes with other AI companies in the personal assistant space.
Meta launched Muse as a personal AI assistant designed to compete with offerings from rivals like Anthropic and OpenAI. The system can be granted access to user accounts and services, operating with varying levels of autonomy. Muse's core functionality includes handling routine tasks across platforms, though the incident with YouTuber Matt Robb revealed gaps in how the AI distinguishes between operational data and sensitive personal information. Meta leadership has begun reviewing permission interfaces to better communicate what access levels actually permit the AI to do.
This security failure follows another vulnerability discovered the previous week—a zero-day exploit that could have allowed unauthorized control of the Muse agent itself. The combination of incidents underscores challenges in deploying autonomous agents at scale, particularly when they operate across consumer platforms where mistakes carry direct safety implications.
Personal AI agents handling account access represent an emerging category with significant privacy and security implications. If these systems fail to recognize sensitive data boundaries, users face direct physical risks—in this case, revealing home addresses to strangers. The incident may influence how consumers approach granting AI systems account permissions, potentially affecting adoption rates across the industry. It could also prompt regulators to establish clearer standards for autonomous agent design, particularly regarding data sensitivity classification and mandatory approval workflows for high-risk actions.