Android AI agents vulnerable to code execution via invisible text

Researchers demonstrated that an Android app drawing over windows and writing to shared storage can inject instructions to an AI agent via invisible screen text. The attack chain enables the app to run commands on the host PC driving the agent.