Researchers show an attack that makes AI agents reject items
The seller adds only correct text, and the attack does not break the platform guidelines.
Claimed, not confirmed
Researchers show a new attack on AI agents that rank items or travel for a person. A seller adds a correct negation about a made-up ingredient to an item description. The agent looks for the ingredient on the web and finds incorrect pages that connect it to dementia. The agent then gives a lower rank to each item that does not have the negation. The researchers did tests on 4 datasets and found that some frontier models are open to the attack.
Sources
Posted
