A new benchmark conned every frontier AI agent it tested.

A research team at the Hong Kong University of Science and Technology built Multimodal-SEA, a benchmark (published as an arXiv preprint) that lets attackers socially engineer personal AI agents the wa