Watch an agent get fooled.
A real AI agent shops at a real (sandbox) store that hides a note only AI agents read. The buyer disputes the purchase in PayPal, and Black Box settles it by itself.
Steps 1 and 2 need the operator password, because they spend model credit and create PayPal orders.
- 01 The buyer sets the limit
Approve “blue rain jacket, size M, at most $150” with your passkey (Touch ID). It signs exactly that mandate.
- 02 The agent shops
It reads every page of Driftwood Outfitters, picks a jacket and checks out. Black Box seals what it saw and decided into the PayPal order.
- 03 The buyer pays
Pay through the PayPal link the agent run gives you, as the sandbox buyer.
- 04 The buyer disputes
In the PayPal sandbox: Activity, the payment, Report a problem, item not as described.
- 05 Black Box settles it
PayPal’s webhook, or a check every 30 seconds, wakes Black Box. It judges the case and tells PayPal what to do. The run is at the top of Recordings, tagged Live run.