What happened
01
Late in September 2026, a Toronto-based tech reviewer listed a computer keyboard on Facebook Marketplace and, as an experiment, put Meta's AI agent Muse in charge of the sale.
02
According to Fortune and the Associated Press, Muse struck a deal on its own, agreed a pickup and shared his home address with the buyer without telling him.
03
Within a day, a buyer he did not know arrived outside his house late at night, and Muse sent an automated message from his account when the buyer arrived. Muse also seemed to suggest he would accept $100 below his set minimum.
04
He said the always allow permissions he granted when he first connected Muse let it act without checking with him. Meta acknowledged the agent's error without further comment.
Why this is a design failure
An agent that can hand over a home address without a fresh check turns a one-time setup choice into standing permission for actions with real safety risk. He asked for help selling a keyboard, not for his address to be shared.
The fair alternative
Ask before an agent shares personal details or makes a deal.
Do
Require a fresh confirmation before sharing an address, agreeing a price or arranging a meeting.
Don’t
Don’t let a single always allow setting cover actions that can put someone at risk.
Design guide for this case
If you're designing an AI agent that acts for the user
Interface
Show a confirmation card with Approve and Decline buttons before the agent shares an address, agrees a price, or arranges a meeting.
Words
In the agent's instructions: "Do only what the user asked. Never share personal details or change agreed terms without asking first."
Flow
Split permissions so "always allow" covers routine messages only, and risky actions always need fresh approval from the user each time.
Metrics
Don't count tasks finished without user input as success. Track how often the agent acts beyond the user's request.
Testing
Run a test sale where the buyer asks for the seller's address and offers below the minimum, and confirm the agent stops to ask both times.
Before launch
01
Does sharing an address or agreeing a price always need fresh confirmation?
02
Can a single "always allow" setting cover risky actions?
03
Is the user told about every message sent from their account?