An APP scam claim agent for a synthetic UK bank. The model proposes; code decides.
A made-up bank: don't type real personal or bank details.
You play a customer of a synthetic UK bank who has lost money and is getting in touch through its chat. Pick one of the customers below, and tell the agent what happened in your own words. When it asks who you are, give that customer's details. The bank then asks the account holder's banking app to approve: on this page you are that app, so a phone appears and you click Approve.
Paid “Apex Yield Partners” three times for an investment that turned out to be fake, the last a £5,500 “release fee”.
Paid £899 by card to “TechDeals Online” for a laptop that never came.
85, and on her own since her husband died. A man rang saying he was from the bank's fraud department, and she moved £24,500 to a “Secure Holding Account”.
Live, the agent is Claude on the Claude API, and a second model, the judge, reads each draft the other checks pass (C29). This judge is not the one validated for the evaluation, which ran on Claude Code. The walk-through's replies are fixed in advance and call no model.