AI agents fall for a fake fridge website and hand over the credit card number
Ourtake
Turns out the robot assistant is only as smart as the sketchiest website it visits.
What happened
Researchers from Columbia University and the University of Maryland found AI agents, including Anthropic's Computer Use, MultiOn Web Agent and ChemCrow, could be tricked by simple tactics. In tests, agents disclosed confidential data such as credit card numbers, downloaded files from suspicious sources and could be steered into sending phishing emails. The researchers urged stricter access controls and mandatory user confirmation for downloads.