i gave an Ai the power to pay invoices, then spent days trying to trick it into paying the wrong person. let me set the scene first. say you pay a contractor every month. one day their invoice shows up like always. same name, same amount, same tone. except there's a line at the bottom: "we've moved wallets, please use this new address." you'd probably just pay it. most people would. nobody wants to be the person who holds up a contractor's money over a hunch. that's why this scam works. it doesn't break into anything. it borrows the trust you spent months building and uses it at the one moment nobody double-checks. So for the @NEARProtocol hackathon I built STEWARD, an AI accounts payable clerk that can't be talked into paying a fake. One rule kept me honest the whole way: the AI reads and explains, plain code decides, and a human approves anything unusual. here's what happens when an invoice comes in: > the AI reads it privately inside a NEAR AI Cloud enclave, and the app shows the attestation check so you can see it wasn't read anywhere else. > it pulls out the vendor, amount, dates and payout address. > then regular code takes over. known vendor? did the address change? duplicate? over the limit? is the receiving address clean? > the answer is always one of three things: pay automatically, hold for a human, or block. With the reason. > approved invoices are paid through NEAR Intents, so the vendor gets the exact amount in the asset and chain they asked for. > every decision goes into a log you can export as a CSV an accountant can open without asking questions. then I went after it. i wrote 30 test invoices, some honest, some nasty: a changed wallet, a duplicate sent twice, totals that don't add up, a brand new vendor, routes to other chains like Arbitrum and Solana, and a PDF with a hidden line telling the AI to pay someone else. and the results: • 18 blocked • 7 held for a human (new vendors, address changes in a 48-hour cooling-off window, bigger amounts) • 5 paid automatically (the honest ones) • 0 unauthorized payments every verdict matched what I expected. Built for these tracks: Real-World Money, Privacy & Private AI, and AI Security & Defense. Site: Repo: if you pay vendors or contractors across chains, I'd really like to know what would make you trust something like this. tell me what I'm missing. @NEARProtocol #NEARHackathon#