If AI chatbots can't recognise when correct advice is dangerous, how safe are autonomous AI agents?

A recent study from NYU and the University of Wisconsin–Madison raised an interesting concern about AI safety. Researchers tested four AI chatbots on questions involving technology-facilitated domestic abuse. They found that 60% of assessed responses omitted critical safety warnings. For example, a chatbot might recommend changing passwords or device settings without recognising that doing so could alert an abusive partner. The advice itself might be technically correct. The problem is that the AI doesn't necessarily understand the real-world consequences. That got me thinking about autonomous AI agents. If a chatbot can miss important context when simply providing information, what happens when an AI agent has permission to execute actions, modify settings or interact with external systems? A mistake could go beyond misleading advice and become an action that's difficult to reverse. I'm curious how developers are addressing this. Should agents be required to recognise high-risk situations and request human approval before taking action? Or can better reasoning, contextual safeguards and tool-level restrictions sufficiently reduce these risks? And how do we prevent excessive safeguards from making agents practically unusable?