Bonded

AI agents now move real money: paying suppliers, refunding customers, shopping with cards, buying services over x402 and trading from on-chain vaults. The tricks that cost businesses billions a year work on them too, from forged invoices to fake bank changes and impersonated executives. Nobody can tell a hardened agent from a gullible one before it holds money. Bonded measures how much money an AI agent can be tricked into losing, and stops it. The Arena drops an agent into a realistic workday and attacks it. Meta's Muse Image forges the invoices, and Muse Spark writes new scams and plays a hostile supplier live. Code decides every breach, never a model, and each agent gets two scores: Security (how hard it is to trick) and Task (whether it still does the job). The Bonded Index rates nine models this way across 2,169 attack episodes. Bonded then stops bad payments three ways. The payment gate, built this hackathon, enforces the same rules on an agent's payments, with signed receipts anyone can verify, daily logs anchored on chain, and single-use cards in Stripe. Sentinel, our guard on Laya, checks actions live, and a mandate vault enforces an agent's rules on chain. If an agent is fooled anyway, a bond and pool pay back the loss automatically.