Approval cards cut forged attack success from 0.54 to 0.03
A 2nd device gave no gain that you can measure when the trusted confirmation is on the same host.
Claimed, not confirmed
Researchers give 12 attacks on the approval step that AI agents use before they send money or erase files. A test of 10 agent runtimes found no approval record that a third party can check. Their reference defense, ASU, shows the user a card that comes from a set description of the task. A passkey on the phone of the user signs the card.
Sources
Posted
