Safety · 2026-10-05

AI agents: new threats to privacy and safety

Google researchers describe open privacy and safety issues for AI agents, that is, programs that can independently carry out tasks on a person's behalf.

✨ What it's about

Google Research has published a piece on open and still unresolved privacy and safety issues for AI agents. This refers to programs built on generative AI that not only answer questions but also take actions themselves: they search for information, work with files or emails, make purchases, or plan tasks.

The title of the research points to a contextual approach. This means risks should be assessed not in isolation from the situation, but with attention to who uses the agent, what data it sees, and what it uses that data for.

📌 Why it matters for ordinary people

The more independence AI gains, the more questions arise: what data an agent can see, whether it might pass that data on to someone it shouldn't, and whether scammers could spoof it. The answers to these questions are not yet final, and the researchers explicitly describe them as open.

For now, it's useful to keep one simple rule in mind: don't give AI apps access to personal data and important accounts beyond what is truly necessary, and check the actions the agent takes on your behalf.

Sources: 🔗 Google Research Blog