Design Real-Time Fraud Detection

Cash App scores every transaction in real time and keeps a running risk rating on each customer: an activation rating set at signup, and a continuous one that moves as the customer behaves. Start the design from that rating. The generic answer is e-commerce card fraud: a stolen card at a checkout, score the purchase, approve or decline. That is not where Block bleeds. Block moves money between people and between sellers and buyers, and the expensive fraud lives in the money-movement path and in the account itself, not at a shopping cart.

Say the threat model out loud in the first minute, because picking it correctly is most of the signal in this round. One fact shapes everything: a lot of Block's money movement is fast and hard to reverse. A Cash App balance sends to another balance instantly, a cash-out pushes to an external debit card over card-network rails, and a Bitcoin withdrawal leaves for good. Once money has cashed out to a mule's debit card or left as Bitcoin, you are not clawing it back the way a card issuer reverses a chargeback. That speed is the product itself, and it is also the attack surface.

So the system watches a stream of transfer, cash-out, login, and account-change events and decides, fast, whether to let a money-moving action through. It is not scoring purchases at a cart. And it has to work across business units that get abused differently. Cash App P2P, Square seller payments, and Afterpay installments each leak money in their own shape, and the feature work below has to cover all three.

Name the four abuse shapes you are defending against, because they drive the features later. P2P scams are the signature one: a real customer is socially engineered into authorizing a send to a scammer, so the account is not compromised, the person is. Account takeover, ATO, is someone logging into a real customer's Cash App, relinking a card, and draining the balance to a new debit card or to Bitcoin.