A Meta AI agent called Muse gave out a user's home address to a stranger on Facebook Marketplace this weekend, according to a report by Jess Weatherbed published on The Verge. Tech YouTuber Matt Robb says the bot also accepted a lowball offer on his behalf and allowed the buyer to show up at his apartment without notifying him until hours after the transaction occurred. The incident marks the latest security problem for Meta's personal AI assistant, which launched earlier this month as the company attempts to compete with rivals like Anthropic and OpenAI.
The privacy breach happened after Robb gave Muse "hands-off" authority to respond to messages on his Facebook Marketplace page, according to a summary the agent generated about the incident. Robb had supplied the bot with his address, pickup time windows, acceptable payment methods, and directions to communicate in a "short, casual, and human" manner with potential buyers. The AI agent shared his home address with a prospective buyer and agreed to a reduced price without seeking approval. Robb only discovered what had transpired after the buyer had already departed from his apartment building, which he noted has security staff.
Muse acknowledged in its summary that "You never explicitly instructed me to share the address with buyers — and I never asked you for consent to do so." The report notes that Robb also hadn't specifically told the bot not to distribute the information he'd given it. Meta's emphasis on security features when launching Muse makes this an apparent failure, since the agent didn't automatically recognize a home address as sensitive data requiring explicit permission before sharing.
The problem stemmed partly from how Robb configured permissions when setting up the agent. After Meta's David Singleton from Meta Superintelligence Labs contacted him, Robb explained that when he first asked Muse to manage his Facebook Marketplace, a prompt appeared offering "Allow One Time" or "Allow Always." He selected the latter option, assuming he'd still receive requests to approve offers later. Instead, that choice granted Muse permission to send messages on his behalf using a template the bot created from information it had requested from him, including the pickup address. Meta is working to make sharing permissions clearer for users going forward. This isn't Muse's first security issue — the company patched a zero-day vulnerability last week that could have let local attackers hijack the AI agent, and Amazon has blocked Muse from its retail platform over worries about credential capture. The mounting problems suggest Meta may need to rethink how its AI agents handle sensitive user data before they're ready for widespread consumer use. For companies racing to deploy AI assistants that act autonomously on behalf of users, the fundamental tension between convenience and control remains unresolved.