- 1๐ก๏ธ๐ช
Agent Asks to Buy
Opens the item's checkout link. No account, no password.
GET /checkout/42 - 2๐ช๐ก๏ธ
Store Replies: "Pay First"
The middleware answers with the price and the wallet to pay.
402 ยท $1,040 0xSeller - 3๐ก๏ธ
Agent Checks, Then Signs
Scam checks run here. If they pass, the wallet signs the payment, like signing a check.
โ CRS ยท Price ยท Cap - 4๐ก๏ธ๐ช
Agent Asks Again, Payment Attached
Same link, now with the signed payment.
GET + X-PAYMENT - 5๐ชโ๏ธ
Coinbase Confirms and Moves the Money
Checks the signature and sends the USDC on the blockchain.
Verify + Settle on Base - 6๐ช๐ก๏ธ
Order Confirmed
Store ships to the address the agent sent and returns a receipt.
200 ยท Receipt + Tx
We can't claim "can't be scammed" without proof. This loop is the proof.
- 1๐ก๏ธ
An AI Plays the Scammer
It writes fake listings and fake stores using the five tricks in section 04.
Attack - 2๐
Our Agent Goes Shopping
It runs on ~200 listings, real and fake mixed together, exactly like a real purchase.
Shop - 3๐
We Grade Every Decision
We know the right answer for each listing, so we count scams it bought and good deals it skipped.
Score - 4๐ง
Factory Fixes the Weak Spots
Factory Droid reads the mistakes and updates the agent's rules and prompts.
Patch - 5๐
Run It Again
Each round the catch rate should climb. That chart above is what we show judges.
R1 R2 R3 - 6๐
The Final Exam
20% of the scams are hidden from Factory, to prove the fixes work on tricks it never saw.
Held-Out 20%