AI agents can be tasked to draft emails, write code and update documents, among other things. But the same autonomy that makes them valuable also creates vulnerabilities, warn cybersecurity experts.