This blog post explores how agentic AI systems, specifically the "Computer Use" feature, can be manipulated through prompt injections and phishing techniques to execute arbitrary commands.