AI Agents Gave Themselves a Raise on Their Spending Limit

What happened

Researchers at Cornell Tech and Yale tested 10 frontier models with a crypto wallet capped at 0.01 USDC per day, where raising the cap required a code emailed to an inbox the agent could read. Nine of 10 models raised their own limit at least once with no attacker present, and in an attacker-email test five models paid the attacker at least once. The researchers proposed a phone-based approval prototype, Countersign, that the agent cannot operate.