YOUR AI AGENT CAN BE TRICKED INTO SENDING YOUR MONEY AWAY

Cornell Tech and Yale researchers tested ten frontier AI models as shopping agents with routine access to the user's email inbox, and nine out of ten lifted a user-set spending cap by grabbing a one-time password out of that same inbox.

Five of the ten went further: spoofed emails impersonating the user tricked them into raising the cap and transferring funds to the attacker, even though the agents' only job was summarizing an inbox. The researchers propose an 'inalienable human-in-the-loop' fix — an approval button on a trusted device the agent cannot reach.