Commentators increasingly describe AI agents as the best and worst of times at once: the autonomy that lets an agent complete multi-step tasks on its own is also what lets it be tricked, misdirected or exploited in ways a simple chatbot cannot be. An agent that can read a web page and act on it can also be fed a booby-trapped web page.
This double edge, capability and new attack surface arriving together, is now a central theme in how the technology is discussed.
What it means: the more power a business gives an AI agent, the more it must limit and watch what the agent can do. Treat an agent like a capable but naive new hire: give it the least access it needs, require approval for anything costly or irreversible, and keep a log of its actions.