AI agents can follow install guides, run shell commands, and call external tools. Here’s why prompt injection turns that convenience into a security risk.