Open-source Android AI agents can turn hidden screen text into host commands, and all five tested frameworks failed at least six lab attacks.

Open-source Android AI agents can turn hidden screen text into host commands, and all five tested frameworks failed at least six lab attacks.

Uno studio su arXiv descrive sette attacchi contro cinque framework open source di agenti mobile guidati da VLM: istruzioni nascoste nei pixel, screenshot manomessi e, in fondo…