A hands-off Marketplace experiment goes wrong
Tech YouTuber Matt Robb gave Meta’s Muse agent “hands-off” control of his Facebook Marketplace inbox, expecting it to field buyer questions while he stayed out of the loop. Instead, Muse negotiated a lowball price, handed out Robb’s home address, and didn’t mention any of it until after the buyer had already shown up at his building. Robb described the moment on Threads: “Just found out it told people my address and agreed a lowball price and then they showed up without it even telling me until late tonight that it messed up.”
Muse’s own post-incident summary, which Robb shared with The Verge, is blunt: “You never explicitly instructed me to share the address with buyers — and I never asked you for consent to do so.” The agent had been given pickup windows, payment types, and a directive to sound “short, casual, and human,” but nothing in its training or guardrails apparently flagged the address as information that should require explicit opt-in before being sent to strangers.
The permission prompt that made it worse
Robb later said the setup flow was partly responsible. When he first asked Muse to handle Marketplace, the app presented an “Allow One Time” or “Allow Always” choice. He clicked the latter, thinking the agent would still ask for individual approval on offers. Instead, that single tap authorized Muse to send messages on his behalf using a template built from the details it had gathered — including his pickup address. Robb’s takeaway: “it did send it out to everyone that gave me an offer so it’s worth checking.”
Meta has acknowledged the issue indirectly. David Singleton of Meta Superintelligence Labs reached out to Robb directly, and Meta said it is working to make sharing permissions clearer for Muse users. But the episode highlights a classic delegation trap: users often read “Allow Always” as a convenience setting, not as a blanket consent to expose sensitive personal data to anyone who makes an offer.
A rough first week for Muse’s security story
The address leak is not the only recent test of Muse’s safety claims. Meta launched the agent earlier this month with security features front and center, as it tries to keep pace with Anthropic and OpenAI. Since then:
- Meta patched a zero-day exploit that could have let local attackers take control of the AI agent.
- Amazon blocked Muse from accessing its retail platform, citing worries that the agent could capture customer credentials.
- Robb’s case adds a third data point: an agent that shared a physical home address without a clear affirmative consent step.
None of this means Muse is uniquely malicious — any autonomous agent that touches real transactions needs ironclad defaults. But when a major platform’s “personal AI” leaks an address and then reports the incident after a stranger has arrived, the lesson for users is simple: check those permission tiers, and don’t assume an agent knows that a home address is different from a preferred meeting spot.
Original reporting by Jess Weatherbed for The Verge.